Application of particle filter algorithm based on chaotic sequences and improved t-distribution in UWB indoor positioning Authors: Bing Li, Xingshao Chai, Shanshan Yang, Jingmei Zhou, Yanxin Hao, Chungang Liu Published: 2026 ABSTRACT Ultra-wideband (UWB) technology has significant application value in indoor autonomous navigation and robot path planning. However, noise interference in complex environments often causes anomalies in UWB positioning data, severely reducing positioning accuracy and stability. A particle filter algorithm (LTPF) based on chaotic sequences and an improved t-distribution is proposed to address this issue. This algorithm applies a linear transformation to the Logistic map to generate chaotic sequences, thereby enhancing particle diversity and distribution. Additionally, it uses standardized residuals to improve the t-distribution and construct an observation likelihood function, thereby reducing the impact of noise points and outlier data. Experimental results show that compared to the PF, EKPF, KGAPF, and SKF-DMM algorithms, the LTPF algorithm achieves average positioning accuracy improvements of 47.80%, 40.53%, 43.73%, and 50.10%, respectively, effectively enhancing positioning accuracy and stability. Additionally, its computational complexity is equivalent to the PF algorithm, making it a viable method for UWB positioning in complex indoor environments. Introduction With the growing demand for high-precision indoor positioning, Ultra-Wideband (UWB) has emerged as the primary technology for indoor positioning due to its wide bandwidth, low power consumption, strong anti-interference capabilities, and centimetre-level positioning accuracy1. However, in practical applications, factors such as non-line-of-sight (NLOS) and multipath effects caused by obstacles in complex environments2, clock drift and synchronization drift in devices3, and height restrictions and unreasonable base station layouts during deployment4 can introduce noise points and abnormal data into UWB positioning data, significantly reducing positioning accuracy and stability5, thereby limiting the application of UWB in fields such as robot indoor autonomous navigation and path planning6. Therefore, researching UWB positioning methods suitable for complex indoor environments is significant. Filtering algorithms7 are commonly used for optimizing positioning data processing. Standard filtering algorithms include the Kalman filter (KF)8 and particle filter (PF)9. Literature10 improved the standard Kalman filter (SKF) based on the global offset method and the measurement discard method, respectively, and compared the positioning performance of the two improved methods with that of SKF. Among them, the improved method based on measurement discard removes abnormal data by setting a threshold, demonstrating superior positioning performance. However, since SKF relies on the Gaussian distribution assumption and linearization approximation, its application in NLOS environments is limited. In contrast, particle filtering uses a Monte Carlo sampling strategy to estimate the posterior distribution of the target state, demonstrating greater adaptability and robustness in handling non-linear and non-Gaussian noise11. In indoor positioning, particle filtering primarily faces the issues of particle degeneration and incorrect weight allocation. Both of these issues can lead to a decrease in the estimation accuracy of particle filtering, thereby reducing positioning accuracy and stability. Particle degeneration occurs because, during the iteration process, a small number of particles obtain excessively high weights. In contrast, the weights of most particles become very small, resulting in a reduction in the number of effective particles and a decrease in particle diversity12. Reference13 analyses the impact of particle likelihood probability distributions on filtering accuracy, proposes a hierarchical particle diversity metric, and sets a threshold for evaluating particle diversity after resampling. Gaussian random walks are employed to enhance diversity if particle diversity fails to meet the threshold. However, this algorithm relies on threshold settings, and improper threshold selection may result in ineffective improvements to particle diversity. Reference14 proposes a particle filtering algorithm based on local linear embedding, which reduces high-dimensional particle states to a low-dimensional Euclidean space while preserving the linear relationships between particles. Particles are then divided into a retention group and an adjustment group based on their weights, with smaller-weighted particles adjusted and mapped back to the high-dimensional space. Finally, the two groups of particles are merged after resampling for the following iteration calculation. However, this algorithm introduces additional parameters, and improper parameter settings may result in poor particle diversity enhancement effects. Reference15 proposes an improved particle filtering resampling method that uses high-weight particles as parent particles and constructs new particles based on a Gaussian distribution to maintain particle diversity. Reference16 combines the advantages of residual resampling and polynomial resampling to assess the degradation level of particles in the current state, thereby selecting the most suitable resampling strategy. Although these two methods reduce the loss of particle diversity compared to traditional resampling methods, they cannot avoid the sample depletion phenomenon caused by multiple copies of high-weight particles. Additionally, these methods typically use Gaussian distributions or other random distributions to generate particle sets, causing particles to cluster in local regions and exacerbating particle degradation. Since particle weights depend on the degree of match between each particle and the observed data, the higher the match, the greater the weight. A significant deviation between the observed data and the actual location can lead to incorrect particles being assigned excessively high weights17. Reference18 uses the Kalman filter algorithm to construct an importance probability density function and introduces a genetic algorithm to perform selection, crossover, and mutation operations on the particle set during the resampling process, making the particle distribution approach the posterior probability distribution. Reference19 uses the extended Kalman filter (EKF) to update the sampled particles and constructs the importance probability density function using the EKF values, making the particle distribution approach the posterior probability distribution. Reference20 uses the firefly algorithm to evaluate the particle distribution to control the optimization intensity and detect the particle density around the optimal particle. When particles are overly aggregated, the elastic mechanism of the spring model is used to adjust the particle distribution, making it more reasonable. Reference21 proposes a particle filtering method based on an improved artificial bee colony algorithm. This algorithm simulates the foraging process of bees to guide particles toward regions with a high likelihood of posterior probability, thereby improving the rationality of the particle distribution. Although the above methods optimize the particle distribution by improving the importance of the probability density function and introducing intelligent optimization algorithms, they still cannot avoid the problem of incorrect weight allocation caused by significant observational data errors. Therefore, to address the issues of particle degeneration and incorrect weight allocation in particle filter algorithms applied to UWB positioning, this paper proposes a UWB positioning method based on LTPF. The main contributions of this paper are as follows:To address the issue of particle degeneration, this paper proposes a method based on logistic mapping to enhance particle diversity. A chaotic sequence with good exploration properties is generated by applying a linear transformation to the Logistic mapping. This sequence is combined with positioning data to initialize the particle set. Additionally, the global error variation is analyzed, and a state transition equation is established based on the chaotic sequence, thereby optimizing the particle distribution while enhancing particle diversity.To address the issue of incorrect weight allocation, a method is proposed to improve the t-distribution-based observation likelihood function using standardized residuals. The observation likelihood function is constructed based on the t-distribution, and standardized residuals are used to analyze the error magnitude of the current observation data. The degrees of freedom of the t-distribution are dynamically adjusted to control its tail thickness, thereby effectively suppressing the influence of noise points and outliers.Combining the improvements from (1) and (2), the LTPF algorithm is designed. Its positioning performance is tested in three simulation environments with different noise intensities and two real-world scenarios with varying environmental complexities. Experimental results show that the LTPF algorithm has the same computational complexity as traditional particle filter algorithms and can improve positioning accuracy and stability in complex indoor environments. To address the issue of particle degeneration, this paper proposes a method based on logistic mapping to enhance particle diversity. A chaotic sequence with good exploration properties is generated by applying a linear transformation to the Logistic mapping. This sequence is combined with positioning data to initialize the particle set. Additionally, the global error variation is analyzed, and a state transition equation is established based on the chaotic sequence, thereby optimizing the particle distribution while enhancing particle diversity. To address the issue of particle degeneration, this paper proposes a method based on logistic mapping to enhance particle diversity. A chaotic sequence with good exploration properties is generated by applying a linear transformation to the Logistic mapping. This sequence is combined with positioning data to initialize the particle set. Additionally, the global error variation is analyzed, and a state transition equation is established based on the chaotic sequence, thereby optimizing the particle distribution while enhancing particle diversity. To address the issue of incorrect weight allocation, a method is proposed to improve the t-distribution-based observation likelihood function using standardized residuals. The observation likelihood function is constructed based on the t-distribution, and standardized residuals are used to analyze the error magnitude of the current observation data. The degrees of freedom of the t-distribution are dynamically adjusted to control its tail thickness, thereby effectively suppressing the influence of noise points and outliers. To address the issue of incorrect weight allocation, a method is proposed to improve the t-distribution-based observation likelihood function using standardized residuals. The observation likelihood function is constructed based on the t-distribution, and standardized residuals are used to analyze the error magnitude of the current observation data. The degrees of freedom of the t-distribution are dynamically adjusted to control its tail thickness, thereby effectively suppressing the influence of noise points and outliers. Combining the improvements from (1) and (2), the LTPF algorithm is designed. Its positioning performance is tested in three simulation environments with different noise intensities and two real-world scenarios with varying environmental complexities. Experimental results show that the LTPF algorithm has the same computational complexity as traditional particle filter algorithms and can improve positioning accuracy and stability in complex indoor environments. Combining the improvements from (1) and (2), the LTPF algorithm is designed. Its positioning performance is tested in three simulation environments with different noise intensities and two real-world scenarios with varying environmental complexities. Experimental results show that the LTPF algorithm has the same computational complexity as traditional particle filter algorithms and can improve positioning accuracy and stability in complex indoor environments. UWB positioning principles Time of arrival positioning technology UWB positioning primarily estimates the distance or angle between a target and an anchor node by measuring time-of-flight (TOF)22, time difference of arrival (TDOA)23, time of arrival (TOA)24, or angle of arrival (AOA)25, thereby determining the target’s position26. Among these, TOA is a positioning technique that determines the target’s position by measuring the time it takes for a signal to travel from the transmitter to the receiver. It calculates the distance between the tag and the base station by multiplying the signal propagation time by the propagation speed. Due to its high-ranging accuracy, TOA is widely used in UWB positioning systems27. Assume that the coordinates of the ℓth base station are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( a_{\ell },b_{\ell },c_{\ell } \right)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell = 1,2,\cdots ,m$$\end{document}, and the coordinates of the tag are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( x,y,z \right)$$\end{document}. Then, the formula for calculating the distance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{\ell }$$\end{document} between the tag and the base station is:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r_{\ell } = \sqrt{\left( a_{\ell } - x \right) ^{2} + \left( b_{\ell } - y \right) ^{2}+ \left( c_{\ell } - z \right) ^{2} } = c\left( t_{\ell }- t_{0} \right) \end{aligned}$$\end{document}where c is the propagation speed of the signal, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{0}$$\end{document} is the time when the tag transmits the signal, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\ell }$$\end{document} is the time when the signal arrives at the base station. Chan algorithm The Chan algorithm is a classic positioning analysis algorithm that solves for the position of the label by performing two weighted least squares solutions on the ranging data28. Compared to other analytical methods such as Taylor29, its advantage lies in its lower computational complexity, enabling it to quickly analyze multiple sets of ranging data while maintaining high accuracy30. According to Eq. (1), we obtain:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r_{1}^{2} - \left( a_{1}- x \right) ^{2} - \left( b_{1}- y \right) ^{2} - \left( c_{1}- z \right) ^{2} = r_{\ell }^{2} - \left( a_{\ell }- x \right) ^{2} - \left( b_{\ell }- y \right) ^{2} - \left( c_{\ell }- z \right) ^{2} \end{aligned}$$\end{document}Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{\ell } = a_{\ell }^{2} + b_{\ell }^{2}+ c_{\ell }^{2}$$\end{document}:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r_{1}^{2}-F_{1}+2a_{1}x+2b_{1}y+2c_{1}z=r_{\ell }^{2}-F_{\ell }+2a_{\ell }x+2b_{\ell }y+2c_{\ell }z \end{aligned}$$\end{document}Convert to matrix form:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r_1^2-r_\ell ^2-F_1+F_\ell =2 \begin{bmatrix} a_{\ell ,1}&b_{\ell ,1}&c_{\ell ,1} \end{bmatrix} \begin{bmatrix} x \\ y \\ z \end{bmatrix} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{\ell ,1}= a_{\ell }- a_{1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{\ell ,1}= b_{\ell }- b_{1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_{\ell ,1}= c_{\ell }- c_{1}$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{a}= \left[ \begin{matrix}x&y&z \end{matrix} \right] ^{T}$$\end{document}, and establish the linear equation:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h= GZ_{a} \end{aligned}$$\end{document}where:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&h= \begin{bmatrix} r_{1}^{2}-r_{2}^{2} - F_{1}+ F_{2} \\ r_{1}^{2}-r_{3}^{2} - F_{1}+ F_{3} \\ \vdots \\ r_{1}^{2}-r_{\ell }^{2} - F_{1}+ F_{\ell } \end{bmatrix} \end{aligned}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&G= 2\begin{bmatrix} a_{2,1} & b_{2,1} & c_{2,1}\\ a_{3,1} & b_{3,1} & c_{3,1}\\ \vdots & \vdots & \vdots \\ a_{\ell ,1} & b_{\ell ,1} & c_{\ell ,1} \end{bmatrix} \end{aligned}$$\end{document}Due to noise interference in the indoor environment, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h\ne GZ_{a}$$\end{document}, assume that the error vector is:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varphi =h-GZ_a=cY\gamma +0.5c^2\gamma \odot \gamma \end{aligned}$$\end{document}where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y= \text {diag} \left( r_{1},r_{2},\cdots ,r_{\ell } \right)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {diag}(\cdot )$$\end{document} denotes a diagonal matrix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = \begin{bmatrix} \gamma _{1}&\gamma _{2}&\cdots&\gamma _{\ell } \end{bmatrix}^{T}$$\end{document} represents the delay error, and ⊙ denotes the Hadamard product. Perform weighted least squares on Eq. (8):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Z_a=\left( G^T\psi ^{-1}G\right) ^{-1}\left( G^T\psi ^{-1}h\right) \end{aligned}$$\end{document}where ψ is the covariance matrix of the error vector φ:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \psi =E\left[ \varphi \varphi ^T\right] =c^2Y\xi Y \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\xi = E\left[ \gamma \gamma ^{T} \right]$$\end{document} is the covariance matrix of TOA. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{a}^{\prime } = \left[ \begin{matrix}x^{2}&y^{2}&z^{2} \end{matrix} \right] ^{T}$$\end{document}, and establish the linear equation:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varphi ^{\prime }=h^{\prime }-G^{\prime }Z_a^{\prime } \end{aligned}$$\end{document}where12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&h^{\prime }= \begin{bmatrix} \left( x+e_1\right) ^2 \\ \left( y+e_2\right) ^2 \\ \left( z+e_3\right) ^2 \end{bmatrix} \end{aligned}$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&G^{\prime }=\begin{bmatrix} 1& 0 & 0 \\ 0 & 1& 0\\ 0 & 0& 1 \end{bmatrix} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e_{1}, e_{2}, e_{3}$$\end{document} are the estimated errors of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{a}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi ^{\prime }$$\end{document} is the updated error vector, and Eq. (11) is solved using weighted least squares:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Z_{a}^{\prime }=\left[ \left( G^{\prime }\right) ^T\left( \psi ^{\prime }\right) ^{-1}G^{\prime }\right] ^{-1}\left[ \left( G^{\prime }\right) ^T\left( \psi ^{\prime }\right) ^{-1}h^{\prime }\right] \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi ^{\prime }$$\end{document} is the covariance matrix of the error vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi ^{\prime }$$\end{document},15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\psi ^{\prime }=E\left[ \varphi ^{\prime }\left( \varphi ^{\prime }\right) ^T\right] =4Y^{\prime }Cov\left[ Z_a\right] Y^{\prime } \end{aligned}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&Cov\left[ Z_a\right] =E\left[ \Delta Z_a\Delta Z_a^T\right] =\left[ G^T\psi ^{-1}G\right] ^{-1} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta Z_a$$\end{document} is the estimation error of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_a$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y^{\prime } = \text {diag}\left( x,y,z \right)$$\end{document}, thus the label position is solved as:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Z= \left[ \pm \sqrt{Z_{a}^{\prime } } \right] \end{aligned}$$\end{document} Improvements to the particle filter algorithm Particle filter algorithm Particle filtering (PF) is a recursive Bayesian estimation algorithm based on the Monte Carlo method. Its core idea is to represent the probability distribution of the target state using a set of random samples (called particles) and recursively update the particles’ weights and states to approximate the actual posterior distribution31. The main steps of the algorithm are as follows:Initialisation: At the initial time k = 0, N particles are randomly sampled from the prior distribution, and each particle is assigned the same weight. Typically, the prior distribution is assumed to follow a Gaussian distribution. The formulas for calculating the particle state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{0}^{i}$$\end{document} and weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{0}^{i}$$\end{document} are: Initialisation: At the initial time k = 0, N particles are randomly sampled from the prior distribution, and each particle is assigned the same weight. Typically, the prior distribution is assumed to follow a Gaussian distribution. The formulas for calculating the particle state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{0}^{i}$$\end{document} and weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{0}^{i}$$\end{document} are: Initialisation: At the initial time k = 0, N particles are randomly sampled from the prior distribution, and each particle is assigned the same weight. Typically, the prior distribution is assumed to follow a Gaussian distribution. The formulas for calculating the particle state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{0}^{i}$$\end{document} and weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{0}^{i}$$\end{document} are: 18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&X _{0}^{i} \sim \mathcal {N}\left( 0,Q \right) ,i=1,2,\cdots ,N \end{aligned}$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\omega _{0}^{i} = \frac{1}{N} \end{aligned}$$\end{document}where Q is the covariance of the process noise.(2)(2) Prediction: At each time step, predict the new position of the particle based on the state transition equation, which is: (2)(2) Prediction: At each time step, predict the new position of the particle based on the state transition equation, which is: (2) Prediction: At each time step, predict the new position of the particle based on the state transition equation, which is: 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} X_{k}^{i} =X_{k-1}^{i} +\chi _{k-1}^{i} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi _{k-1}^{i} \sim \mathcal {N}\left( 0,Q \right)$$\end{document} is the perturbation noise.(3)Update weight: Update the weight of each particle based on the new observation data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{k}$$\end{document}, with the weight update formula being: (3)Update weight: Update the weight of each particle based on the new observation data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{k}$$\end{document}, with the weight update formula being: Update weight: Update the weight of each particle based on the new observation data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{k}$$\end{document}, with the weight update formula being: 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \omega _{k}^{i}=\omega _{k-1}^{i}p\left( Z_{k}\mid X_{k}^{i}\right) ,i=1,2,\cdots ,N \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p\left( Z_{k}\mid X_{k}^{i}\right)$$\end{document} is the observation likelihood function, which typically assumes that the observation noise follows a Gaussian distribution. Its calculation formula is:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p\left( Z_k\mid X_k^i\right) =\frac{1}{\sqrt{2\pi R}}\exp \left( -\frac{\left( Z_k-X_k^i\right) }{2R}\right) \end{aligned}$$\end{document}where R is the covariance of the observation noise. After updating the weights of all particles, normalization is performed:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \omega _{k}^{i}= \frac{\omega _{k}^{i}}{ {\textstyle \sum _{i= 1}^{N}}\omega _{k}^{i} } \end{aligned}$$\end{document}(4)Resampling: Since some particles’ weights may become very small during the update process while others’ weights remain relatively large, resampling is needed to avoid particle degeneration. N particles are randomly selected from the current particle set based on their weights, forming a new particle set, and the particle weights are reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{0}^{i} = \frac{1}{N}.$$\end{document} (4)Resampling: Since some particles’ weights may become very small during the update process while others’ weights remain relatively large, resampling is needed to avoid particle degeneration. N particles are randomly selected from the current particle set based on their weights, forming a new particle set, and the particle weights are reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{0}^{i} = \frac{1}{N}.$$\end{document} Resampling: Since some particles’ weights may become very small during the update process while others’ weights remain relatively large, resampling is needed to avoid particle degeneration. N particles are randomly selected from the current particle set based on their weights, forming a new particle set, and the particle weights are reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{0}^{i} = \frac{1}{N}.$$\end{document} Improvements to particle set initialization and state transition equation Traditional particle filtering algorithms typically assume that particles are sampled and propagated according to a Gaussian distribution, causing most particles to cluster around the mean and accelerating particle degradation, which makes it difficult to meet UWB positioning requirements in complex environments. Therefore, this paper employs a logistic mapping to enhance the particle set’s state space coverage. Its mathematical expression is:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{k+1}=\alpha q_{k}\left( 1-q_{k}\right) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \ \in (0,4]$$\end{document} is a control parameter that determines the system’s dynamical behavior. When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \ \in (0,3],$$\end{document} the system gradually converges to a stable point; when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \ \in (3,3.57)$$\end{document}, the system enters a state of period doubling; and when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \ \in [3.57,4]$$\end{document}, the system exhibits chaotic behavior, characterized by sensitivity to initial conditions and long-term unpredictability. To evaluate the sensitivity of the proposed method to the parameter α, experiments were conducted under conditions where the noise standard deviation was 0.5 m. As shown in Fig. 1, the root-mean-square positioning error remains relatively stable within the chaotic region, indicating that the proposed method is not sensitive to the specific value of α. When α approaches the lower bound of the region, particle diversity decreases slightly, leading to a small increase in positioning error. Based on the observed performance trend, α is set to 3.77 in this study.Fig. 1Relationship between particle number and positioning accuracy. Relationship between particle number and positioning accuracy. Logistic mapping can generate chaotic sequences with good randomness and employability. However, since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{k}$$\end{document} is restricted to the narrow range (0,1), it limits its application in positioning. To address this issue, this paper applies a linear transformation to the Logistic mapping expression, thereby generating chaotic sequences within the range (-1,1):25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} q_{k+1}=-\frac{\alpha }{2} \left( q_{k}^2-1\right) -1 \end{aligned}$$\end{document}Figure 2 shows the particle set initialized using a Gaussian distribution and the particle set initialized using the Logistic map with the linear transformation. It can be observed that the particle distribution generated by the Gaussian distribution is more concentrated. In contrast, the particle set generated by the Logistic map with linear transformation exhibits higher diversity and more uniform distribution characteristics.Fig. 2Comparison of particle set initialization methods. Comparison of particle set initialization methods. When initializing the particle set, the observed data at time k = 1 is used to generate the particle set, thereby reducing the search space and improving the initial accuracy of the particle set:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} X_{k}^{i} =Z_{1}+\xi _{k+1}^{i},k=0 \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_{1} =\begin{bmatrix} x_{1}&y_{1}&z_{1} \end{bmatrix} ^{T}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\xi _{k+1}^{i} =\begin{bmatrix} q_{x,k+1}^i&q_{y,k+1}^i&q_{z,k+1}^i \end{bmatrix} ^{T}$$\end{document}denotes the chaotic perturbation vector in the three spatial directions, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{x,k+1}^i$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{y,k+1}^i$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{z,k+1}^i$$\end{document} are independently generated chaotic variables for the x, y, and z axes, respectively. When setting the state transition equation, calculate the mean error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta _k$$\end{document} between the true estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{Z}_k$$\end{document} and the observed data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z_k$$\end{document}, and combine it with the chaotic sequence to set the disturbance noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _k^i$$\end{document}:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\delta _{k}=\frac{1}{k} \sum _{j=1}^{k} \left| \hat{Z} _{j}-Z_{j} \right| \end{aligned}$$\end{document}28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\theta _{k}^{i} =q_{k+1}^{i} \delta _{k} \chi _{k}^{i} \end{aligned}$$\end{document}The state transition equation is:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} X_{k}^i=X_{k-1}^i+\theta _{k-1}^i \end{aligned}$$\end{document} Improvements in observing likelihood function The observation likelihood function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p\left( Z_{k}\mid X_{k}^{i}\right)$$\end{document} of the particle filter algorithm is used to calculate the weight of each particle. However, non-Gaussian noise and abnormal observation data easily affect the traditional observation likelihood function, leading to incorrect particle weight calculations and reduced positioning accuracy. Therefore, to enhance the robustness of particle filtering against noise, this paper utilizes the heavy-tailed characteristics of the t-distribution and standardized residuals to construct the observation likelihood function. In addition, the proposed method only modifies the observation likelihood model, while the particle propagation, importance sampling, and resampling framework of the standard particle filter remain unchanged. According to the convergence theory of Sequential Monte Carlo methods proposed by Pierre Del Moral, particle filter estimation converges to the true posterior distribution when the importance weights are non-negative, bounded, and properly normalized32. Since the adopted t-distribution likelihood function is continuous, non-negative, and integrable, the modified observation model still satisfies the convergence conditions of the standard particle filter framework. Therefore, the proposed LTPF algorithm preserves the convergence properties of conventional particle filtering methods, with the calculation formula being:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p\left( Z_k\mid X_k^i\right) =\frac{\Gamma \left( \frac{\zeta _{k} +d}{2}\right) }{\Gamma \left( \frac{\zeta _{k} }{2}\right) (\zeta _{k} \pi )^{\frac{d}{2}}\left| R\right| ^{\frac{1}{2}}}\left( 1+\frac{D_k^i}{\zeta _{k} }\right) ^{-\frac{\zeta _{k} +d}{2}} \end{aligned}$$\end{document}where d is the dimension of the observation vector, for three-dimensional positioning d = 3, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Gamma \left( \cdot \right)$$\end{document} denotes the gamma function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_{i}^{k}$$\end{document} is the Mahalanobis distance between the particle and the observation data:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} D_k^i= \begin{pmatrix} Z_k-X_k^i \end{pmatrix}^TR^{-1} \begin{pmatrix} Z_k-X_k^i \end{pmatrix} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\zeta _{k}$$\end{document} is the degrees of freedom, which controls the tail thickness of the t-distribution. In this paper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\zeta _{k}$$\end{document} is set as the standard deviation residual, with the specific steps as follows: First, define the observed data residual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{k}$$\end{document} as the Euclidean distance between the observed data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z _{k}$$\end{document} and the previous moment’s true estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{Z} _{k-1}$$\end{document},32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \beta _{k}=\left\| Z_{k} -\hat{Z} _{k-1} \right\| \end{aligned}$$\end{document}Second, calculate the standard deviation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _k$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{k}$$\end{document} to measure the dispersion of errors at time k:33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \sigma _k=\sqrt{\frac{1}{k}\sum _{j=1}^k\left( \beta _j-\overline{\beta }\right) ^2} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\beta }$$\end{document} is the mean residual up to time k:34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \bar{\beta } = \frac{1}{k} \sum _{j= 1}^{k}\beta _{j} \end{aligned}$$\end{document}Finally, To avoid numerical instability caused by excessively small variance values during the early filtering stage or near-perfect estimation conditions, a small positive constant \varepsilon is introduced into the denominator of the standardized residual calculation. set the degree of freedom \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\zeta _{k}$$\end{document} to:35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \zeta _{k} =\frac{\beta _{k} }{\sigma _{k }+ \varepsilon } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon = 10^{-6}$$\end{document}. Resampling step This paper adopts the systematic resampling method, whose core idea is to reduce randomness by introducing a fixed offset. This reduces the variance in the resampling process and prevents a decrease in the number of effective particles. First, define the cumulative weight of the first i particles as:36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C_k^i=\sum _{j=1}^i\omega _k^j \end{aligned}$$\end{document}Next, uniformly generate N sampling points in the interval \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ 0,1 \right]$$\end{document}:37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} u_i=\frac{i-1+\textrm{u}}{N} \end{aligned}$$\end{document}where u∼ U(0,1) is a random number following a uniform distribution. Determine the index \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J_{k}^{i}$$\end{document} of the new particle based on the cumulative weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{k}^{i}$$\end{document}:38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} J_{k}^{i} =\min \left\{ j|u_i\le C_k^j\right\} \end{aligned}$$\end{document}Then, update the particle set and set the same weight for the new particles:39\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&X_k^i=X_k^{J_k^i} \end{aligned}$$\end{document}40\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\omega _k^i=\frac{1}{N} \end{aligned}$$\end{document}Finally, calculate the true estimate using the weighted average method:41\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{Z}_k=\sum _{i=1}^N\omega _k^iX_k^i \end{aligned}$$\end{document} Algorithm complexity analysis The computational complexity of the LTPF algorithm is primarily influenced by the number of particles N and the time step k. During the initialization phase, N particles are generated based on a chaotic sequence and assigned uniform weights, with a computational complexity of O(N); during the prediction phase, particle states are updated using state transition equations, with a computational complexity of O(N); in the update phase, the particle weights are updated by constructing an observation likelihood function based on the t-distribution and standardized residuals, followed by normalization, with a computational complexity of O(N); finally, in the resampling phase, the particle set is updated using a systematic resampling method, and the final estimate is calculated using a weighted average method, with a computational complexity of O(N). The overall complexity of the LTPF algorithm is:42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} O\left( k\cdot N \right) \end{aligned}$$\end{document}Therefore, the LTPF algorithm effectively improves positioning accuracy and stability while maintaining the same complexity as traditional particle filter algorithms. Algorithm steps The detailed process of the UWB positioning method based on LTPF is shown in Algorithm 1, and the specific steps can be summarised as follows: Set algorithm parameters: Set the number of particles to N.Data collection: Collect TOA data between tags and base stations, calculate the tag’s initial position using the Chan algorithm, and use this data as the observation data for the LTPF algorithm.Initialisation step: Generate a particle set using a chaotic sequence and the observation data at k=1, and assign equal weights to all particles.Prediction step: Combine the chaotic sequence and the mean of the overall error to set the perturbation noise, establish the state transition equation, and predict the particle positions.Weight update step: Construct an observation likelihood function based on the t-distribution, set the degrees of freedom as the standardized residual between the current observation data and the previous accurate estimate, then calculate and normalize the particle weights.Resampling step: Update the particle set using the system resampling method, assign the same weights, and calculate the true estimated position of the tag using the weighted average method.Iterate the algorithm by repeating steps 4 through 6. The algorithm terminates once all T observations have been processed. Set algorithm parameters: Set the number of particles to N. Set algorithm parameters: Set the number of particles to N. Data collection: Collect TOA data between tags and base stations, calculate the tag’s initial position using the Chan algorithm, and use this data as the observation data for the LTPF algorithm. Data collection: Collect TOA data between tags and base stations, calculate the tag’s initial position using the Chan algorithm, and use this data as the observation data for the LTPF algorithm. Initialisation step: Generate a particle set using a chaotic sequence and the observation data at k=1, and assign equal weights to all particles. Initialisation step: Generate a particle set using a chaotic sequence and the observation data at k=1, and assign equal weights to all particles. Prediction step: Combine the chaotic sequence and the mean of the overall error to set the perturbation noise, establish the state transition equation, and predict the particle positions. Prediction step: Combine the chaotic sequence and the mean of the overall error to set the perturbation noise, establish the state transition equation, and predict the particle positions. Weight update step: Construct an observation likelihood function based on the t-distribution, set the degrees of freedom as the standardized residual between the current observation data and the previous accurate estimate, then calculate and normalize the particle weights. Weight update step: Construct an observation likelihood function based on the t-distribution, set the degrees of freedom as the standardized residual between the current observation data and the previous accurate estimate, then calculate and normalize the particle weights. Resampling step: Update the particle set using the system resampling method, assign the same weights, and calculate the true estimated position of the tag using the weighted average method. Resampling step: Update the particle set using the system resampling method, assign the same weights, and calculate the true estimated position of the tag using the weighted average method. Iterate the algorithm by repeating steps 4 through 6. The algorithm terminates once all T observations have been processed. Iterate the algorithm by repeating steps 4 through 6. The algorithm terminates once all T observations have been processed. Algorithm 1UWB positioning method based on LTPF UWB positioning method based on LTPF Simulation experiments and analysis To validate the performance of the proposed algorithm, a simulation environment with varying degrees of noise was constructed using the MATLAB R2022b platform, and experimental analyses were conducted on the PF31, EKPF19, KGAPF18, SKF-DMM10, and LTPF algorithms. The experiments used a Dell laptop computer with an Intel(R) Core(TM) i5-9300H CPU @ 2.40 GHz and 8 GB of RAM. Algorithm parameter settings In particle filter algorithms, the number of particles N is a key parameter that controls algorithm performance, directly affecting the algorithm’s estimation accuracy and computational complexity. In a simulation environment with noise standard deviation ranging from 0 to 1m, the relationship between different numbers of particles and positioning accuracy is shown in Fig. 3. As can be seen, as the number of particles increases, the root mean square error (RMSE) of the PF, EKPF, KGAPF, and LTPF algorithms gradually decreases. When the number of particles exceeds 600, the RMSE stabilizes, and even if the number of particles is further increased, the RMSE does not decrease significantly.Fig. 3Relationship between particle number and positioning accuracy. Relationship between particle number and positioning accuracy. Table 1 shows the relationship between the number of particles and the algorithm runtime for 300 data sets. It can be seen that as the number of particles increases, the algorithm runtime also increases. This is because the KGAPF algorithm incorporates the Kalman filter and genetic algorithms. In contrast, the EKPF algorithm incorporates the extended Kalman filter algorithm, both of which introduce additional computational overhead, resulting in longer runtime. The runtime of the LTPF algorithm is similar to that of the PF algorithm, with both running times being less than 3 seconds when the number of particles is 600. Therefore, this paper sets the number of particles to 600 to ensure positioning accuracy while reducing time consumption.Table 1Comparison of algorithm runtime under different particle numbers (unit:s).NumberPFEKPFKGAPFLTPF1000.430.671.310.472000.791.242.290.834001.412.093.951.586002.163.246.092.288002.934.348.223.1410003.835.5910.863.9812004.436.4812.584.16 Table 1Comparison of algorithm runtime under different particle numbers (unit:s).NumberPFEKPFKGAPFLTPF1000.430.671.310.472000.791.242.290.834001.412.093.951.586002.163.246.092.288002.934.348.223.1410003.835.5910.863.9812004.436.4812.584.16 Comparison of algorithm runtime under different particle numbers (unit:s). Algorithm positioning performance comparison To verify the positioning performance of the algorithm in different noise environments, three simulation environments with noise standard deviations of 0.2 m, 0.5 m, and 0.8 m were constructed in a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$20m\times 20m\times 20m$$\end{document} space. The LTPF algorithm, the PF algorithm, the Improved Particle Filter (ISPF) with modified particle set initialization and state transition equations, and the Improved Likelihood Function Particle Filter (ILPF). As shown in Fig. 4, the ISPF algorithm constrains the particle propagation process by improving the generation method of particle sets and the state transition model, thereby suppressing unreasonable and abrupt state transitions, resulting in a noticeable shift of the curve to the left. The ILPF algorithm adapts to outlier data by refining the observation likelihood function; its cumulative distribution function (CDF) curve is shifted to the left compared to that of the PF algorithm, mitigating the impact of outlier observations. The LTPF algorithm combines the ISPF and ILPF algorithms, maintaining a relatively forward-shifted and smooth CDF curve under all three noise conditions. Its cumulative probability exhibits more stable behavior as error increases, indicating that this method can effectively reduce positioning error.Fig. 4Comparison of CDF Curves. Comparison of CDF Curves. To further validate the localization performance of the LTPF algorithm, we conducted experimental comparisons between the LTPF, PF, EKPF, KGAPF, and SKF-DMM algorithms. Figure 5 shows the CDF of the error for different algorithms under noise standard deviations of 0.2 m, 0.5 m, and 0.8 m, respectively.Fig. 5CDF plots of the algorithm under three noise standard deviations. CDF plots of the algorithm under three noise standard deviations. As can be seen, as the noise standard deviation increases from 0.2 to 0.8 m, the CDF curves of all methods exhibit varying degrees of rightward shift overall, with the error distribution range significantly expanding. This indicates that elevated noise levels exert a pronounced impact on positioning accuracy. At a noise standard deviation of 0.2 m, all algorithms exhibit rapid accumulation trends within the small-to-medium error range, indicating relative advantages at low error levels. As the noise standard deviation increases to 0.5 m and 0.8 m, differences between methods gradually emerge: the CDF curves for the PF and SKF-DMM methods shift markedly to the right, with increased dispersion in error distribution. In contrast, the LTPF method maintained relatively advanced and smooth CDF curves under all three noise conditions. Its cumulative probability exhibited more stable variation with increasing error, indicating superior error suppression capability under varying noise levels. To provide a more intuitive comparison of the overall performance of each algorithm, Table 2 presents the RMSE and standard deviation (SD) of each algorithm under different noise intensities, which are used to measure the accuracy and stability of the algorithms, respectively. In addition, the 95% confidence intervals (CI) of the RMSE are also reported in the table to further reflect the statistical reliability of the results. Figure 6 shows the box plots of the positioning errors for each algorithm. Table 3 shows the corresponding algorithm parameter settings. As can be seen from this, the SKF-DMM method exhibits insufficient robustness and significant positioning errors when encountering noise interference, as it cannot set an appropriate threshold to eliminate outlier data. Although the EKPF method employs extended Kalman filtering to address nonlinear errors and achieves lower RMSE and Std compared to the KGAPF method, it still fails to effectively mitigate the impact of outlier data. In contrast, the LTPF method suppresses noise interference by adjusting the state transition model and observation likelihood function to control particle propagation states and weight distribution. It maintains low RMSE and SD across all three noise environments, with relatively gradual error growth, demonstrating excellent positioning performance. Based on the results of the three sets of experiments, compared to the PF, EKPF, KGAPF, and SKF-DMM algorithms, the LTPF algorithm achieves average improvements in positioning accuracy of 49.09%, 42.38%, 45.68%, and 52.80%, respectively, and average improvements in stability of 53.60%, 47.61%, 49.59%, and 55.50%, respectively.Table 2Comparison of algorithm performance under three noise standard deviations (unit:m).Algorithm0.20.50.8RMSE (95% CI)SDRMSE (95% CI)SDRMSE (95% CI)SDPF0.151 [0.145, 0.157]0.0570.463 [0.445, 0.481]0.1570.696 [0.669, 0.723]0.236EKPF0.137 [0.131, 0.143]0.0510.405 [0.389, 0.421]0.1390.601 [0.578, 0.624]0.207KGAPF0.147 [0.141, 0.153]0.0530.431 [0.414, 0.448]0.1470.627 [0.603, 0.651]0.212SKF-DMM0.164 [0.157, 0.171]0.0590.513 [0.494, 0.532]0.1700.727 [0.700, 0.754]0.240LTPF0.091 [0.088, 0.094]0.0270.206 [0.198, 0.214]0.0670.334 [0.321, 0.347]0.116 Table 2Comparison of algorithm performance under three noise standard deviations (unit:m).Algorithm0.20.50.8RMSE (95% CI)SDRMSE (95% CI)SDRMSE (95% CI)SDPF0.151 [0.145, 0.157]0.0570.463 [0.445, 0.481]0.1570.696 [0.669, 0.723]0.236EKPF0.137 [0.131, 0.143]0.0510.405 [0.389, 0.421]0.1390.601 [0.578, 0.624]0.207KGAPF0.147 [0.141, 0.153]0.0530.431 [0.414, 0.448]0.1470.627 [0.603, 0.651]0.212SKF-DMM0.164 [0.157, 0.171]0.0590.513 [0.494, 0.532]0.1700.727 [0.700, 0.754]0.240LTPF0.091 [0.088, 0.094]0.0270.206 [0.198, 0.214]0.0670.334 [0.321, 0.347]0.116 Comparison of algorithm performance under three noise standard deviations (unit:m). Fig. 6Box plots of the algorithm under three noise standard deviations. Table 3Settings for each algorithm.ItemPFEKPFKGAPFLTPFParticle InitializationGaussian initializationGaussian initializationGaussian initializationLogistic map initializationImportance Density FunctionGaussian distributionConstructed based on EKFConstructed based on KFImproved state transition equationLikelihood ModelGaussian likelihoodGaussian likelihoodGaussian likelihoodt-distribution likelihood with standardized residualResampling StrategySystematic resamplingSystematic resamplingGenetic algorithm improved resamplingSystematic resampling Box plots of the algorithm under three noise standard deviations. Table 3Settings for each algorithm.ItemPFEKPFKGAPFLTPFParticle InitializationGaussian initializationGaussian initializationGaussian initializationLogistic map initializationImportance Density FunctionGaussian distributionConstructed based on EKFConstructed based on KFImproved state transition equationLikelihood ModelGaussian likelihoodGaussian likelihoodGaussian likelihoodt-distribution likelihood with standardized residualResampling StrategySystematic resamplingSystematic resamplingGenetic algorithm improved resamplingSystematic resampling Settings for each algorithm. Actual experiments and analysis To further validate the positioning accuracy and stability of the LTPF algorithm, this paper conducts positioning experiments in two scenarios: an indoor repair workshop and an indoor warehouse. The results are compared and analyzed with those of the Chan28, PF, EKPF, KGAPF, and SKF-DMM algorithms. Experimental environment The experiment was set up with five base stations and one tag using the DW1000 UWB chip produced by DECAWAVE. The positioning module’s main control chip is the STM32F103 series MCU, with a ranging error of less than 0.1m. During positioning, the device transmits data at a rate of approximately 2.5 messages per second. The computer communicates with the tag via a serial port to collect movement data along a predefined trajectory. Simultaneously, a laser rangefinder is used to measure the actual positions of the base stations and tags, with a ranging accuracy of up to ± 5mm.Fig. 7Indoor repair workshop base station layout diagram.Fig. 8Indoor repair workshop scene.Table 4Base station locations in the indoor repair workshop (unit:m).Base stationLocation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 0,0,0 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 1.032,7.446,-0.287 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 6.871,6.581,-1.319 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 7.789,-2.720,-1.124 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 0.985,-4.05,-0.296 \right)$$\end{document} Indoor repair workshop base station layout diagram. Indoor repair workshop scene. Table 4Base station locations in the indoor repair workshop (unit:m).Base stationLocation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 0,0,0 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 1.032,7.446,-0.287 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 6.871,6.581,-1.319 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 7.789,-2.720,-1.124 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 0.985,-4.05,-0.296 \right)$$\end{document} Base station locations in the indoor repair workshop (unit:m). The dimensions of the indoor repair workshop scene are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$11.61m\times 9.27m\times 3.72m$$\end{document}. The floor and surrounding walls are constructed of cement and metal materials. The scene includes seven motor vehicles, one car lift, and various metal equipment and tools. Base stations are deployed on walls and tripods. The coordinates of the base stations are shown in Table 4, and their deployment locations are shown in Fig. 7. Red squares represent base stations, and pedestrians represent tags. The actual scene is shown in Fig. 8. The indoor warehouse measures \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.3m\times 7.3m\times 3.7m$$\end{document} in length, width, and height, respectively. The metal shelves measure \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6m\times 0.5m\times 2m$$\end{document} and are divided into three layers, each with a height of 0.85 m. The shelves store metal instruments and equipment, packaging boxes, mechanical parts, and other items. The base stations are installed on the walls around the perimeter of the indoor space. The coordinates of the base stations are shown in Table 5, the installation locations are shown in Fig. 9, and the actual scene is shown in Fig. 10.Table 5Base station locations in the indoor warehouse (unit:m).Base stationLocation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 0,0,0 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 1.805,3.75,-2.217 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 7.68,3.75,-0.187 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 7.55,-3.33,-0.351 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 1.82,-3.18,-2.217 \right)$$\end{document} Table 5Base station locations in the indoor warehouse (unit:m).Base stationLocation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{1}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 0,0,0 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 1.805,3.75,-2.217 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 7.68,3.75,-0.187 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 7.55,-3.33,-0.351 \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS_{5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( 1.82,-3.18,-2.217 \right)$$\end{document} Base station locations in the indoor warehouse (unit:m). Fig. 9Indoor warehouse base station layout diagram. Fig. 10Indoor warehouse scene. Indoor warehouse base station layout diagram. Indoor warehouse scene. Experimental analysis and verification Figure 11 compares the trajectories calculated by different algorithms and the actual path in an indoor repair workshop. The blue curve represents the true trajectory, while the red curves denote the estimated trajectories from each algorithm. The performance of the algorithms can be evaluated by assessing the closeness of the estimated trajectories to the true trajectory. As shown in the figure, the Chan algorithm exhibits significant fluctuations in its positioning results throughout the process, deviating considerably from the true trajectory. This indicates its limited accuracy and susceptibility to interference from outlier data. Although the PF method, EKPF method, and KGAPF method improve trajectory smoothness to some extent, local deviations still occur. The SKF-DMM algorithm exhibits even more pronounced trajectory deviations, indicating its limited ability to suppress interference from complex environments. In contrast, the LTPF method demonstrates high consistency between its positioning results and the actual trajectory. Its trajectory exhibits smoother morphology and more concentrated spatial distribution, effectively mitigating the impact of anomalous data on positioning outcomes.Fig. 11Algorithm trajectory map in an indoor repair workshop scene. Algorithm trajectory map in an indoor repair workshop scene. Figure 12 shows the CDF of positioning errors for different algorithms in an indoor repair workshop. The figure indicates that the CDF curve of the Chan algorithm grows slowly, suggesting poor positioning accuracy and significant environmental interference. In contrast, the PF, EKPF, KGAPF, SKF-DMM, and LTPF algorithms significantly improve the Chan algorithm. Among these, the SKF-DMM algorithm’s curve remains within 0.3m for most cases, and the EKPF and KGAPF algorithms’ curves are mostly within 0.2 m, demonstrating stable error suppression capabilities. In contrast, the EKPF algorithm’s curve is more forward-leaning, showing a sharp increase in the low-error range, proving its reliability in low-interference environments. In contrast, over 90% of the LTPF algorithm’s curves are within 0.1 m, indicating its superior positioning performance compared to other algorithms.Fig. 12CDF plot of the algorithm in an indoor repair workshop scene. CDF plot of the algorithm in an indoor repair workshop scene. Table 6 shows the RMSE and SD for each algorithm in an indoor repair workshop. Figure 13 shows the corresponding box plot. The experimental results indicate that the Chan algorithm exhibits significant errors in this environment, reflecting its lack of robustness when dealing with nonlinear errors in complex environments, making it unsuitable for high-precision positioning requirements. While the EKPF algorithm improves positioning accuracy and stability compared to the PF and KGAPF algorithms, it fails to adapt effectively to environmental changes. In contrast, the LTPF algorithm achieves an RMSE and SD of only 0.116m and 0.037m, significantly outperforming other algorithms. Compared to the Chan, PF, EKPF, KGAPF, and SKF-DMM algorithms, with positioning accuracy improvements of 69.63%, 49.12%, 39.27%, 43.96%, and 51.46%, respectively, and stability improvements of 70.63%, 47.89%, 39.34%, 42.19%, and 54.32%, respectively.Table 6RMSE and SD of the algorithm in an indoor workshop scene (unit:m).AlgorithmRMSE (95% CI)SDChan0.382 [0.368, 0.396]0.126PF0.228 [0.220, 0.236]0.071EKPF0.191 [0.184, 0.198]0.061KGAPF0.207 [0.200, 0.214]0.064SKF-DMM0.239 [0.230, 0.248]0.081LTPF0.116 [0.112, 0.120]0.037 Table 6RMSE and SD of the algorithm in an indoor workshop scene (unit:m).AlgorithmRMSE (95% CI)SDChan0.382 [0.368, 0.396]0.126PF0.228 [0.220, 0.236]0.071EKPF0.191 [0.184, 0.198]0.061KGAPF0.207 [0.200, 0.214]0.064SKF-DMM0.239 [0.230, 0.248]0.081LTPF0.116 [0.112, 0.120]0.037 RMSE and SD of the algorithm in an indoor workshop scene (unit:m). Fig. 13Box plot of the algorithm in an indoor repair workshop scene. Box plot of the algorithm in an indoor repair workshop scene. Figure 14 compares the trajectories calculated by different algorithms and the actual path in an indoor warehouse. As shown in the figure, the overall fluctuation range of positioning results from various algorithms significantly increases in this scenario, with noticeable deviations occurring within local areas. Particularly in path turning regions, error accumulation becomes more pronounced. The Chan algorithm exhibits severe deviation in three-dimensional space. While the PF method, EKPF method, and KGAPF method mitigate error propagation to some extent, their positioning results still display noticeable jitter. In contrast, the LTPF algorithm maintains high spatial consistency with the true trajectory even under enhanced interference conditions. Its positioning results exhibit low dispersion, demonstrating excellent positioning stability and error suppression capability.Fig. 14Algorithm trajectory map in an indoor warehouse scene. Algorithm trajectory map in an indoor warehouse scene. Figure 15 shows the cumulative distribution function of positioning errors for different algorithms in an indoor warehouse. Compared to the indoor repair shop scenario, the Chan algorithm exhibits higher positioning errors, with a slow-growing curve and poorer performance. While the SKF-DMM algorithm outperforms the Chan algorithm, as the error range expands, the curve rises more gradually, and its ability to maintain accuracy significantly weakens. The curves for the EKPF and KGAPF algorithms overlap, with their curves slightly ahead of the PF algorithm curve. Most curves remain within 0.4m, but the error rate significantly increases, exposing their limitations in complex scenarios. The LTPF algorithm curve rises rapidly and approaches one at approximately 0.3m, indicating its effectiveness in suppressing errors.Fig. 15CDF plot of the algorithm in an indoor warehouse scene. CDF plot of the algorithm in an indoor warehouse scene. Table 7 shows the RMSE and SD for each algorithm in an indoor warehouse. Figure 16 shows the corresponding box plot. Overall, this scenario has more obstacles than the indoor repair shop scenario, with more severe signal blocking, multipath effects, and non-line-of-sight propagation. As a result, the RMSE and SD for each algorithm increased in this scenario. However, the LTPF algorithm still maintained the best performance. Compared to the Chan, PF, EKPF, KGAPF, and SKF-DMM algorithms, with improvements in positioning accuracy of 60.63%, 45.18%, 39.95%, 41.56%, and 46.05%, respectively, and improvements in stability of 64.50%, 42.66%, 35.43%, 38.81%, and 44.59%, respectively. This indicates that the algorithm is adaptable and robust to environmental changes.Fig. 16Box plot of the algorithm in an indoor repair workshop scene.Table 7Algorithm in an indoor warehouse scene RMSE and SD (unit:m).AlgorithmRMSE (95% CI)SDChan0.607 [0.585, 0.629]0.231PF0.436 [0.420, 0.452]0.143EKPF0.398 [0.383, 0.413]0.127KGAPF0.409 [0.394, 0.424]0.134SKF-DMM0.443 [0.426, 0.460]0.148LTPF0.239 [0.230, 0.248]0.082 Box plot of the algorithm in an indoor repair workshop scene. Table 7Algorithm in an indoor warehouse scene RMSE and SD (unit:m).AlgorithmRMSE (95% CI)SDChan0.607 [0.585, 0.629]0.231PF0.436 [0.420, 0.452]0.143EKPF0.398 [0.383, 0.413]0.127KGAPF0.409 [0.394, 0.424]0.134SKF-DMM0.443 [0.426, 0.460]0.148LTPF0.239 [0.230, 0.248]0.082 Algorithm in an indoor warehouse scene RMSE and SD (unit:m). Conclusion UWB positioning in complex indoor environments is susceptible to NLOS, multipath effects, clock drift, and improper base station deployment, leading to significant deviations in positioning data and reducing positioning accuracy and stability. Therefore, this paper proposes a UWB positioning method based on LTPF. First, the importance of the particle filter algorithm for optimizing UWB positioning data processing is analyzed. Then, corresponding solutions are designed to address the issues of particle degradation and incorrect weight allocation in the application of the particle filter algorithm for positioning. Finally, simulations and actual experiments demonstrate that the LTPF method can significantly improve UWB positioning accuracy and stability. However, there are still some shortcomings in practical applications. For example, positioning accuracy is fundamentally constrained by the inherent ranging uncertainty of UWB hardware. In practical systems, the ranging error typically fluctuates around 0.1 m under ideal line-of-sight (LOS) conditions, which implies that the achievable positioning accuracy has a theoretical lower bound determined by the statistical characteristics of the measurement noise rather than a fixed constant value. Therefore, the proposed LTPF algorithm improves positioning performance by suppressing abnormal observations, instead of eliminating the inherent measurement uncertainty of the hardware. Furthermore, this paper does not provide a systematic analysis of the effects of the number of base stations, their spatial distribution, and their geometric structure. In addition, the current design of the positioning algorithm does not incorporate dynamic factors such as the time dimension and direction of motion. Future work will address these issues to further improve the performance of the positioning system.