Development of Advanced Positioning Techniques of UWB/Wi-Fi RTT Ranging for Personal Mobility Applications Authors: Harris Perakis, Vassilis Gikas, Günther Retscher Published: 2024 ABSTRACT “Smart” devices, such as contemporary smartphones and PDAs (Personal Digital Assistance), play a significant role in our daily live, be it for navigation or location-based services (LBSs). In this paper, the use of Ultra-Wide Band (UWB) and Wireless Fidelity (Wi-Fi) based on RTT (Round-Trip Time) measurements is investigated for pedestrian user localization. For this purpose, several scenarios are designed either using real observation or simulated data. In addition, the localization of user groups within a neighborhood based on collaborative navigation (CP) is investigated and analyzed. An analysis of the performance of these techniques for ranging the positioning estimation using different fusion algorithms is assessed. The methodology applied for CP leverages the hybrid nature of the range measurements obtained by UWB and Wi-Fi RTT systems. The proposed approach stands out due to its originality in two main aspects: (1) it focuses on developing and evaluating suitable models for correcting range errors in RF-based TWR (Two-Way Ranging) technologies, and (2) it emphasizes the development of a robust CP engine for groups of pedestrians. The results obtained demonstrate that a performance improvement with respect to position trueness for UWB and Wi-Fi RTT cases of the order of 74% and 54%, respectively, is achieved due to the integration of these techniques. The proposed localization algorithm based on a P2I/P2P (Peer-to-Infrastructure/Peer-to-Peer) configuration provides a potential improvement in position trueness up to 10% for continuous anchor availability, i.e., UWB known nodes or Wi-Fi access points (APs). Its full potential is evident for short-duration events of complete anchor loss (P2P-only), where an improvement of up to 53% in position trueness is achieved. Overall, the performance metrics estimated based on the extensive evaluation campaigns demonstrate the effectiveness of the proposed methodologies. 1. Introduction The tracking of vehicles, pedestrians, and assets in any platform mode is crucial for supporting various societal functions, spanning from personal mobility services to safety and well-being. With the rise in affordable “smart” devices, the provision of high-quality positioning services has become essential for a wide range of wirelessly connected devices. While Global Navigation Satellite System (GNSS) technologies offer a global solution of satisfactory accuracy outdoors, their performance degrades significantly in hybrid environments and becomes impractical indoors [1,2]. This paper serves two primary objectives. Firstly, it focuses on developing a framework for characterizing and modeling RF-based ranging observables derived from disparate radio localization technologies using empirical models. Secondly, it aims to design a methodology for collaboratively localizing groups of autonomously moving nodes (pedestrians) indoors. This is achieved using data derived between rover units and on-site fixed devices (i.e., Wi-Fi access points) as well as among neighboring rovers. This is accompanied by the development and implementation of associated assessment procedures for performance evaluation. The paper evaluates the performance of radio localization technologies through controlled experimental trials. Statistical analysis enables the characterization of range observable errors generated by highly accurate Ultra-wide Band (UWB) and less accurate Wi-Fi Round-Trip Time (Wi-Fi RTT) technologies. Analysis offers valuable insights into each technology’s capabilities at various operating conditions (i.e., static or dynamic) and in the presence of obstacles commonly found indoors (walls, pedestrian shadowing, etc.). Furthermore, the paper proposes alternative techniques for range error mitigation, including the development and evaluation of suitable empirical correction (linear and spatial) models. The methodology applied for the collaborative localization of a group of users (e.g., pedestrians) leverages the hybrid nature of the range measurements obtained by UWB and Wi-Fi RTT (Round-Trip Time) systems. Firstly, the proposed algorithm calculates the standalone position of the moving nodes using the existing communication infrastructure (Wi-Fi RTT) for Pedestrian-to-Infrastructure (P2I) ranges. Subsequently, the localization solution obtained in the P2I step is combined with ad hoc UWB Pedestrian-to-Pedestrian (P2P) ranges and orientation observables (i.e., loosely coupled filtering scheme) to serve as a baseline for implementing a distributed collaborative position solution. The collaborative positioning system yields an improvement in accuracy and availability compared to the absolute P2I localization solution, while reducing the need for extensive infrastructure equipment. Position availability refers to the percentage of time during for which the positioning terminal delivers a position solution. The quality metrics of the localization algorithms are assessed with field as well as simulated datasets generated using in-house developed software. The paper is structured as follows: After the introduction in Section 1, Section 2 presents, after the general aspects in Section 2.1, the research questions and objectives and the methodology in Section 2.2 and Section 2.3, respectively. Section 3 then introduces the range-based CP system, where firstly in Section 3.1, the range errors and their mitigation are identified followed by a description of the architectures and algorithms for CP in Section 3.2. In the following, Section 4 deals with the introduction of the different employed range correction models, including the correction model’s development steps in Section 4.1, the detailed discussion of these models in Section 4.2, the orientation-assisted range correction models in Section 4.3, and then their validation in Section 4.4. Section 5 then analyzes the position computation algorithms with the correction models adopted for the UWB and Wi-Fi RTT range observable internal accuracy in Section 5.1 and the Kalman Filter formulation for distributed CP in Section 5.2. Section 6, on data collection and error mitigation, includes Section 6.1, Section 6.2 and Section 6.3 on the description of the test data and equipment employed as well as the properties of the indoor field test campaigns and the range error mitigation for both UWB in and Wi-Fi RTT measurements. Although measurements in outdoor environments have been carried out in this study, the paper focuses more on the results obtained indoors in Section 7, where the position solution estimation is thoroughly discussed in Section 7.1 and Section 7.2 with the localization solutions obtained using field and simulated data, respectively. Finally, Section 8 concludes the paper and provides a detailed outlook on future work which will be carried out by the authors of this contribution. 2. Motivation of This Research 2.1. General Aspects The motivation for this research work stems directly from the recent advances in IT (Information Technology) and MEMS (Micro-Electromechanical Systems) technologies and their global adoption in contemporary smartphone and PDA (Personal Digital Assistant) devices. In this era, the potential of recently introduced RF ranging technologies is significant, especially when integrated with inertial sensor data. The driver for undertaking this research study is the recent advances and potential in the following open research areas:RF (Radio Frequency) ranging technologies of high accuracy in the UWB spectrum have recently been introduced to commercial smartphone devices. The continuously decreasing size and cost of UWB augurs their wide adoption by the industry in the near future; it is forecasted to grow from USD 1.82 billion in 2024 to USD 4.08 billion by 2029 [3].The recently introduced Wi-Fi RTT technology can easily offer P2I ranging services of medium nominal accuracy (0.5 m) as part of the default web access functionality in a widespread and seamless manner [4]. Its fast adoption by the smartphone industry [5] dictates the pressing need for in-depth studying of the ranging capabilities and limitations, as well as developing methods towards improving the derived localization solution.The RF-based TW-ToF (Two-Way Time of Flight) ranging principle introduces inherent and condition-specific (bound to varying devices and environment conditions) inaccuracies of a variable nature [6,7,8,9]. As a result, the need for thoroughly studying its error budget, and the development of alternative error mitigation techniques capable of adapting in different environments, is clearly evident.The combination of UWB and Wi-Fi RTT technologies may offer increased coverage and flexibility of indoor positioning solutions by combining pre-existing P2I ranging infrastructure with ad hoc P2P ranging [10]. This ubiquitous ranging setup utilizing these complementary technologies has still not been extensively studied [11].The goal in real-life applications leans towards reducing the number of infrastructure (anchor) nodes in order to lower procurement and maintenance costs [12,13,14] further enhanced by approaches utilizing Reconfigurable Intelligent Surfaces (RISs) or metasurfaces [15,16]. Therefore, the proposed positioning solutions should incorporate flexible architectures utilizing optimally both the available anchor points and moving nodes. Therefore, a study addressing the alternative setups of available anchors, moving nodes, and their geometry distribution in selected operational scenarios is expected to provide useful insight for designing indoor positioning system (IPS) installations.A crucial problem to be addressed in P2P range-based collaborative decentralized positioning architectures is the mitigation of errors induced due to unknown correlations among the communicating nodes. Relevant studies that have concluded on stable Position Velocity and Timing (PVT) solutions are employing node classification concepts—for instance, using primary nodes equipped with multiple sensors and absolute position knowledge in order to provide internodal positioning input to secondary P2P-only rovers [17,18,19]. This approach requires a multilevel rovers design, whereas the overall system operates with reduced flexibility due to its dependence on the continuous operation of high-cost rovers. Therefore, as long as the mitigation of the propagated position errors for collaborative P2P range-only positioning algorithms is not addressed, the provision of a robust and scalable real-time solution still remains an open issue. RF (Radio Frequency) ranging technologies of high accuracy in the UWB spectrum have recently been introduced to commercial smartphone devices. The continuously decreasing size and cost of UWB augurs their wide adoption by the industry in the near future; it is forecasted to grow from USD 1.82 billion in 2024 to USD 4.08 billion by 2029 [3]. RF (Radio Frequency) ranging technologies of high accuracy in the UWB spectrum have recently been introduced to commercial smartphone devices. The continuously decreasing size and cost of UWB augurs their wide adoption by the industry in the near future; it is forecasted to grow from USD 1.82 billion in 2024 to USD 4.08 billion by 2029 [3]. The recently introduced Wi-Fi RTT technology can easily offer P2I ranging services of medium nominal accuracy (0.5 m) as part of the default web access functionality in a widespread and seamless manner [4]. Its fast adoption by the smartphone industry [5] dictates the pressing need for in-depth studying of the ranging capabilities and limitations, as well as developing methods towards improving the derived localization solution. The recently introduced Wi-Fi RTT technology can easily offer P2I ranging services of medium nominal accuracy (0.5 m) as part of the default web access functionality in a widespread and seamless manner [4]. Its fast adoption by the smartphone industry [5] dictates the pressing need for in-depth studying of the ranging capabilities and limitations, as well as developing methods towards improving the derived localization solution. The RF-based TW-ToF (Two-Way Time of Flight) ranging principle introduces inherent and condition-specific (bound to varying devices and environment conditions) inaccuracies of a variable nature [6,7,8,9]. As a result, the need for thoroughly studying its error budget, and the development of alternative error mitigation techniques capable of adapting in different environments, is clearly evident. The RF-based TW-ToF (Two-Way Time of Flight) ranging principle introduces inherent and condition-specific (bound to varying devices and environment conditions) inaccuracies of a variable nature [6,7,8,9]. As a result, the need for thoroughly studying its error budget, and the development of alternative error mitigation techniques capable of adapting in different environments, is clearly evident. The combination of UWB and Wi-Fi RTT technologies may offer increased coverage and flexibility of indoor positioning solutions by combining pre-existing P2I ranging infrastructure with ad hoc P2P ranging [10]. This ubiquitous ranging setup utilizing these complementary technologies has still not been extensively studied [11]. The combination of UWB and Wi-Fi RTT technologies may offer increased coverage and flexibility of indoor positioning solutions by combining pre-existing P2I ranging infrastructure with ad hoc P2P ranging [10]. This ubiquitous ranging setup utilizing these complementary technologies has still not been extensively studied [11]. The goal in real-life applications leans towards reducing the number of infrastructure (anchor) nodes in order to lower procurement and maintenance costs [12,13,14] further enhanced by approaches utilizing Reconfigurable Intelligent Surfaces (RISs) or metasurfaces [15,16]. Therefore, the proposed positioning solutions should incorporate flexible architectures utilizing optimally both the available anchor points and moving nodes. Therefore, a study addressing the alternative setups of available anchors, moving nodes, and their geometry distribution in selected operational scenarios is expected to provide useful insight for designing indoor positioning system (IPS) installations. The goal in real-life applications leans towards reducing the number of infrastructure (anchor) nodes in order to lower procurement and maintenance costs [12,13,14] further enhanced by approaches utilizing Reconfigurable Intelligent Surfaces (RISs) or metasurfaces [15,16]. Therefore, the proposed positioning solutions should incorporate flexible architectures utilizing optimally both the available anchor points and moving nodes. Therefore, a study addressing the alternative setups of available anchors, moving nodes, and their geometry distribution in selected operational scenarios is expected to provide useful insight for designing indoor positioning system (IPS) installations. A crucial problem to be addressed in P2P range-based collaborative decentralized positioning architectures is the mitigation of errors induced due to unknown correlations among the communicating nodes. Relevant studies that have concluded on stable Position Velocity and Timing (PVT) solutions are employing node classification concepts—for instance, using primary nodes equipped with multiple sensors and absolute position knowledge in order to provide internodal positioning input to secondary P2P-only rovers [17,18,19]. This approach requires a multilevel rovers design, whereas the overall system operates with reduced flexibility due to its dependence on the continuous operation of high-cost rovers. Therefore, as long as the mitigation of the propagated position errors for collaborative P2P range-only positioning algorithms is not addressed, the provision of a robust and scalable real-time solution still remains an open issue. A crucial problem to be addressed in P2P range-based collaborative decentralized positioning architectures is the mitigation of errors induced due to unknown correlations among the communicating nodes. Relevant studies that have concluded on stable Position Velocity and Timing (PVT) solutions are employing node classification concepts—for instance, using primary nodes equipped with multiple sensors and absolute position knowledge in order to provide internodal positioning input to secondary P2P-only rovers [17,18,19]. This approach requires a multilevel rovers design, whereas the overall system operates with reduced flexibility due to its dependence on the continuous operation of high-cost rovers. Therefore, as long as the mitigation of the propagated position errors for collaborative P2P range-only positioning algorithms is not addressed, the provision of a robust and scalable real-time solution still remains an open issue. 2.2. Research Questions and Objectives The main research questions to be answered, and the primary and secondary objectives of this research can be summarized as follows:Primary Objective 1: To develop and test a methodology for identifying and mitigating errors in TWR (Two-Way Ranging) RF range observations;Primary Objective 2: To develop and test a robust, RF range-based positioning approach for groups of pedestrians walking in dynamic environments, considering the hybrid nature of TWR measurements;Secondary Objective 1: To establish and implement a unified Quality Control (QC) framework for the assessment of the correctness and efficiency of the proposed solutions. Primary Objective 1: To develop and test a methodology for identifying and mitigating errors in TWR (Two-Way Ranging) RF range observations; Primary Objective 1: To develop and test a methodology for identifying and mitigating errors in TWR (Two-Way Ranging) RF range observations; Primary Objective 2: To develop and test a robust, RF range-based positioning approach for groups of pedestrians walking in dynamic environments, considering the hybrid nature of TWR measurements; Primary Objective 2: To develop and test a robust, RF range-based positioning approach for groups of pedestrians walking in dynamic environments, considering the hybrid nature of TWR measurements; Secondary Objective 1: To establish and implement a unified Quality Control (QC) framework for the assessment of the correctness and efficiency of the proposed solutions. Secondary Objective 1: To establish and implement a unified Quality Control (QC) framework for the assessment of the correctness and efficiency of the proposed solutions. For the first research objective, the initial sub-objective involves conducting methodical field tests in controlled environments to examine TWR RF range errors. These tests aim to retrieve valuable feedback using actual data. This part of the investigation aims at a thorough statistical analysis and characterization of the nature of errors in range observations produced using the UWB and Wi-Fi RTT technologies. This investigation is expected to provide useful insight concerning the technologies’ capabilities in varying operational conditions (e.g., static or kinematic operation) and against obstacles interaction (e.g., through-wall operation and pedestrian shadowing) encountered commonly indoors. Based on the findings of range error characterization, alternative range error mitigation techniques are suggested. This sub-objective aims at developing, implementing, and evaluating suitable range error correction procedures based on specific empirical range error correction models. These models are built to be valid in environments with common characteristics (i.e., different building layouts and/or materials), data availability setup (i.e., transceivers number and installation geometry), obstacles (i.e., pedestrians and walls), as well as different RF technologies (i.e., UWB and/or Wi-Fi RTT) in order to propose a solution providing robust range error correction. The second topic is addressed: Firstly, the proposed algorithm computes the standalone position of the moving nodes in question aided by the existing communication infrastructure. For this purpose, appropriate localization algorithms have been studied for handling the Pedestrian-to-Infrastructure (P2I) ranges obtained by Wi-Fi RTT. The proposed localization engine implements the range-error correction techniques for mitigating inherent Wi-Fi RTT inaccuracies while providing real-time functionality. The second part of the localization approach integrates the range measurements along with the location of the neighbor moving nodes obtained in the previous step as a basis for the development of the distributed collaborative positioning (CP) solution. This goal resides in ad hoc Pedestrian-to-Pedestrian (P2P) UWB ranging technologies in order to support and extend the positioning availability provided by the P2I absolute pedestrian positioning. Overly, the CP engine should be also capable of integrating range error mitigation models whilst ensuring its ability to operate in real time. Finally, in order for the proposed CP engine to be adaptive enough to compensate for dynamic conditions, it is designed to optimally use the P2I and P2P range information. The main goal is to develop a robust algorithm able to accept a different number of anchors (P2I) while minimizing the effect of propagated P2P correlation-induced positioning errors. Appropriate algorithms are proposed for handling the correlated and uncorrelated errors among communicating inter-ranging moving nodes. It is noted that the proposed CP algorithm adheres to inertial measurements via low-complexity range/IMU fusion in order to compensate for the short-term P2I range unavailability. The first sub-objective refers to the evaluation of the proposed range error correction models under varying conditions, for different datasets collected using both TWR technologies under consideration. The field-testing procedures are repeatable and are able to be performed in varying locations. Through the implementation of extended, dedicated field tests, a detailed analysis and evaluation of the different correction models should lead to concrete proposals suitable for each RF ranging technology. The second sub-objective refers to the detailed and extended testing and assessment of the proposed suite of positioning algorithms using real and simulated data. The design and implementation of dedicated field experiments enable the acquisition of complete datasets for testing the proposed algorithm and for fine-tuning the subsequent, extensive simulations tests. Regarding simulation testing, the development of a ranging and orientation data generator based on simulated or real trajectories enables the evaluation in a controlled and repeatable manner. The generated data simulate the quality characteristics of the TWR technologies in order to provide a straightforward assessment based on controlled reference data. 2.3. Research Methodology The research methodology followed in this work consists of three distinct but inter-related implementation steps:Range measurements calibration/correction phase;Positioning algorithms development;Quality Control of the positioning engine utilizing real and simulated datasets. Range measurements calibration/correction phase; Range measurements calibration/correction phase; Positioning algorithms development; Positioning algorithms development; Quality Control of the positioning engine utilizing real and simulated datasets. Quality Control of the positioning engine utilizing real and simulated datasets. The first key item includes (1) the pre-analysis stage, (2) the correction models development, (3) the error mitigation and models validation, and the (4) kinematic range error correction. Initially in step (1), the raw RF range observables need to undergo a preliminary statistical analysis in order to characterize the TWR technologies’ behavior and pave the road for the following analysis steps. In the correction models’ development, the raw data collected in the previous stage undergo a refinement process. The statistical metrics obtained are used in order to guide data grouping, outlier identification, data exclusion, or even the repetition of data collection campaigns. Specifically, the empirical range error models produce both radial (1D) and spatial (2D) ranges before the implementation on real data. Additionally, the models are combined with available RSS (Received Signal Strength) indicators as well as user orientation information enabling investigation of the environmental effects. For the error mitigation and models validation, the developed ranging errors models are applied on real datasets for static conditions, providing initial feedback regarding model performance. In order to ensure unbiased estimation, the calibration evaluation is performed on data collected for validation purposes and not on the initial data collected for the generation of error models. Moreover, the evaluated range error models are implemented on data referring to well-defined operational conditions (e.g., room type and geometry), providing the evaluation of the error models based on real reference data. And finally for step 1, the kinematic range error correction is applied. Ultimately, the range errors mitigation process is of importance for kinematic positioning sessions. The dynamic characteristics of kinematic data provide the most demanding conditions for TWR range error model validation due to varying environmental effects. For validating the reliability and robustness of the models, the performance evaluation is performed utilizing sets of test trajectories that cover the majority of the available testbed areas. The second key item of the positioning algorithms development includes (1) the tuning of the positioning filter, (2) the collaborative positioning algorithms, and (3) the cross-correlation effect mitigation. For the tuning of the positioning filter, the positioning engine’s core is the Kalman Filter (KF), which is developed and optimized to handle UWB and Wi-Fi RTT raw observables. During this development process, both process and measurement noise is fine-tuned in an optimal manner. The tuning is based on the statistical characterization performed for both TWR technologies in the previous development step. Azimuth information obtained by onboard IMU is included within the KF in order to compensate for epochs of gross ranging errors or ranging measurables unavailability. Then, the collaborative positioning algorithm is developed for providing localization in a robust, scalable, and self-contained manner. For that purpose, a distributed architecture is selected. Considering that the foundation of the system relies on the utilization of Wi-Fi RTT for P2I ranging and UWB for P2P ranging, the mathematical models and algorithms should be designed and developed accordingly. Operational elements for each technology, such as sampling rate, data formatting, and communications scheduling are taken into account. At this stage, the collaborative positioning operation is limited to one rover utilizing multiple ranges from neighbor anchors and rovers, omitting the unknown correlation effects on positioning errors during multiple rovers’ localization. For the cross-correlation effect mitigation, an extension of the developed CP algorithms is proposed based on the approach of the Split Covariance Intersection Filter (SCIF), which can be implemented as a variation of KF. The SCIF approach incorporates the cross-correlation in errors occurring due to the relative measurements among collaborating nodes. The proposed approach utilizes range-only relative measurements and the communicated position state of the rover. As this is a range-based approach, the developed filter also incorporates an adaptive KF feature for compensating for an abrupt orientation change. Since absolute positioning is provided by the P2I ranging technology, the maximum expected positioning performance is bound by the ranging quality of Wi-Fi RTT observables. The third key item of the Quality Control (QC) of the positioning engine utilizing real and simulated datasets includes (1) the field testing campaigns, (2) the raw observables simulator, and (3) the performance evaluation. The selected testbed areas in the field testing campaigns need careful selection in order to ensure typical environmental conditions that are required at the evaluation stage of the proposed approach. For the evaluation, a raw software simulator for ranges and heading is designed and developed to aid in the development, testing, and field-testing design of the CP algorithms. The software is based on the configurable simulated trajectory data of multiple simultaneously roving nodes. Modular errors can be introduced for the range measurements in order to simulate the different quality of the TWR technologies at hand. Moreover, the feature of dynamic anchor availability enables the study of simulated obstruction effects commonly present in indoor environments. The performance evaluation statistics are estimated utilizing the preconfigured reference trajectories. In the performance evaluation, the overall assessment of the proposed approach is performed assuming standard quality metrics for relevant PVT-reliant applications (e.g., trueness, accuracy, availability, and Dilution Of Precision) in varying operational conditions. The estimation of the statistical measures takes place for varying range error correction models, leading to an overall evaluation of the proposed range error mitigation approaches. Finally, the effects of varying motion characteristics of the rovers are examined, as different dynamics directly affect the localization engine performance. 3. Range-Based Collaborative Positioning (CP) This section provides a brief background summary on the basic techniques, the measuring principles, and mathematical fundamentals for indoor CP determination using range observables. Inter-nodal ranging may refer both to range measurements originating from roving nodes to static anchors as well as between roving nodes. A description of the usable positioning techniques and methods as well as optimization algorithms in positioning and sensor fusion is omitted, as it can easily be found in the literature, e.g., in the review paper of one of the authors of this contribution [20]. 3.1. Range Error Identification and Mitigation In order for a positioning technique to produce an optimal solution, it is important that the raw observables (ranges, directions, etc.) have undergone exhaustive preprocessing to mitigate gross and systematic errors; see [21]. In particular, in the indoor environment, which is characterized by NLOS (Non-Line of Sight) conditions and severe signal multipath, the raw range observables can be of low quality. Extensive research is currently being undertaken by many research groups worldwide studying the nature of RF-based range errors and modeling their behavior, aiming at minimizing their effect on the final position solution [22,23]. Moreover, the combined effects of NLOS conditions, multipath, signal attenuation and scattering in wireless positioning systems further deteriorate the position quality, as it is subjected to travel from the transmitter to the receiver through multiple paths. Considering that NLOS conditions represent a major challenge for indoor RF-based positioning applications, various research efforts have focused on methodologies aiming at mitigating the NLOS effects. As in real-life applications, the existence and severity of NLOS conditions is a priori unknown; a research approach should aim at characterizing signals as LOS or NLOS. Thereby, if a signal is identified as a LOS one, then no prior action is required, contrarily to signals detected as NLOS ones. The latter undergo dedicated preprocessing techniques for mitigating the respective errors [24,25]. The distinction between LOS and NLOS observables can rely either on sequential range estimation for outliers’ thresholding or on channel statistics [26]. Relevant studies suggest that the non-Gaussian distribution nature indicates an obstacle when working with KF algorithms since they assume that the measurement errors follow a Gaussian distribution [27,28]. Subsequently, for the indoor cases of a mainly non-Gaussian TWR observation nature, it is expected that they are prone to position quality instability due to model assumptions. Attempts to overcome this limitation usually rely on the adoption of non-linear measurement error models, leading usually to particle filters (PFs) [29,30]. However, a PF solution asks for increased computational complexity, which is not easy to support by handheld, low-cost indoor positioning systems. Alternative approaches include realizations of hybrid KF implementations based on pseudo-position measurements that can handle non-Gaussian error models [31]. While they offer reduced computational complexity compared to PF, they still require increased processing power compared to traditional KFs. An alternative approach for handling the non-linear nature of the range error observables indoors is via a Gaussian Mixture (GM) filter type. Such filters can handle error distributions with multipeaks [32,33]. In effect, they apply multiple Gaussian models to approximate the complex nature of the transmitted signals; however, it is crucial to identify and use the optimal number of Gaussian components to avoid unnecessary computational complexity. While this approach offers increased positioning accuracy for highly noisy measurements, its computational complexity increases dramatically for multinode, range-based positioning. It is noted that while the KF approaches reach their limit in highly non-linear cases, still the EKF offers a viable alternative when handling moderately non-linear error models due to their computationally efficient architecture [34,35]. Empirical RF range error models, on the other hand, rely on the systematic collection of real range observables to extract meaningful statistics that adequately describe their nature and extract range variation behavior that might be encountered during real-life localization applications. Examples of empirical modeling of RF-signal for localization include the approach introduced by Li et al. [6] that relies on an asymmetric, double exponential ranging error distribution model. The error model is formulated through fitting real data, whereas an extension of further tuning the suggested model using range-based parameters is proposed. In the work of Jing et al. [36], a Ranging Quality Indicator (RQI) is established based on UWB signal characteristics paired with the corresponding ranging error used to train a machine learning (ML) algorithm. In this approach, the algorithm produces a set of RQI values in real time and dynamically assigns weights to the range measurements in a UWB/IMU particle filter. In a study by Koppanyi and Toth [37], the original UWB range histograms are found to present multiple peaks attributed to multipath effects. To this effect, a Maximum Likelihood Estimator (MLE) is used for selecting the ranges with the highest probability of true values based on a comparison against the lateration-derived coordinates. Moreover, other empirical error models use range- and position-dependent corrections produced using curve-fitting approaches on real data as illustrated in Figure 1. In a research study by Toth et al. [38], range error calibration is implemented based on a grid of calibration points used for the generation of an ad hoc model. In this approach, the calibration values are used for the 2D linear interpolation forming the calibration function. In the work of Ledergerber and D’Andrea [7], a Sparse Pseudo-input Gaussian Process is trained using the known relative antenna pose (angle) and the error computed using the fixed distances between UWB nodes. The objective is to build an error prediction model that will be utilized in Kalman Filter-based UWB positioning. Regarding the field-testing setups followed for UWB range error analysis and identification, different approaches exist depending on the testing scope. On one hand, when extensive characterization of the experimental area needs to be conducted, the tests are focused on the collection of extensive datasets for the dedicated site. For instance, Li et al. [6] perform a series of static indoor field tests for estimating the range error values using multiple anchor nodes and mobiles. The area’s concrete and steel walls result in predominantly NLoS conditions, while the entire sets of data (LOS, multipath and NLOS) are analyzed together simultaneously with the error models generation for improving positioning performance. On the other hand, for generalization purposes, it is common practice for experimental implementation to take place on test sites featuring different characteristics. In the work of Toth et al. [38], tests are conducted under various observation conditions—i.e., a combined outdoor open area, a forest environment, and indoors. The different environment conditions indicate the varying effects on UWB positioning. Subsequently, the error calibration process is based on known calibration points forming a grid. 3.2. Collaborative Positioning An increased interest towards the development of CP approaches is apparent in the recent literature; nevertheless, the concept is not a new one [39,40]. The increased motivation for CP stems both from the technological developments for utilizing optimally Peer-to-Peer (P2P) communication as well as from the need for minimizing the costs of the permanently installed infrastructure (i.e., anchor RF transceivers) used by traditional RF-based positioning systems. In many cases, P2P communication between nodes is based on technologies that can also offer relative ranging such as Wi-Fi, UWB, and Bluetooth [41,42]. In this regard, CP implementations make use of these technologies both for application-specific data transmission as well as for supporting localization needs. This section presents a short description of CP approaches, their architecture, and the most prevailing CP algorithms, as well as an overview of the implemented CP approaches with varying operational conditions. The network architecture of a CP system can either be a centralized or distributed one [43]. These two different architectures are illustrated in Figure 2. In a centralized architecture [6,17,44,45], as the name suggests, the positions estimation is performed centrally by a localization engine typically located at a control center that collects data from all the remote nodes. Central processing translates at increased processing power considering that state (position, orientation, and velocity) computation of all nodes in the network is undertaken by a single processing engine. Naturally, as the information from all nodes in the network needs to be transmitted to the central unit for the estimation to be complete, this approach also leads to increased communication requirements. In addition, as CP systems rely on a single, central engine processing unit with finite processing and communication capabilities, the expansion for increasing (scalability) the supported number of nodes faces crucial limitations. Notwithstanding an appropriately designed and implemented centralized CP engine offering high-accuracy pose estimation for all nodes and inter-nodal state correlations, it suffers decreased robustness. The dependence on a single, central processing engine for continuous operation results in high-probability operational malfunctions. On the other hand, distributed CP architectures depend on their ability to self-estimate nodal positions based on the measurements and information collected within the CP network [19,36,46]. Practically, in order to achieve this goal, each node in the network needs to be equipped with a portable processing unit and a certain communications infrastructure. This translates to decreased processing capabilities and therefore more stringent limitations on the amount of received data that can be supported, and by extension, the accuracy capabilities of the overall system. A useful trade-off is the ability to operate with limited communication among the collaborating nodes as well as to easily integrate additional collaborating nodes, resulting in a highly scalable system. Perhaps the most crucial weaknesses of the distributed CP approach are their inability to maintain inter-nodal correlation at the network level leading to decrease the mitigation of inter-dependent errors. Table 1 summarizes the strengths and weaknesses of the centralized and distributed CP architectures. As the overall motive of this study on utilizing collaborative localization relies on the ability of independent mobile nodes to handle independent positioning information and relative measurements from neighboring nodes, the interest of the current research is focused on the application of distributed localization. The distributed collaborative localization problem for nodes performing relative range measurements can be addressed using four main positioning algorithms: (1) the non-linear Kalman Filter; (2) the particle filter; (3) the belief propagation; and (4) the covariance intersection. As the first two are well known, in the following, only the third and fourth algorithms are briefly elaborated on. Belief propagation algorithms rely on factor graphs, and particularly on the well-known Sum Product Algorithm over Wireless Network (SPAWN), which is an inherently cooperative localization approach. It relies on the exchange of messages for each node in the network to determine its a posteriori distribution given all the available measurements; see [47]. Despite being able to provide highly accurate results, the SPAWN algorithm also suffers high computational complexity as well as requiring a specialized configuration for handling loopy networks (i.e., the “outbound” ranging observable affects the consequent “inbound” observables) [48,49]. When fusing information among neighboring nodes within a CP network, a highly challenging task is the mitigation of accumulated inter-dependent errors, that is, the computation of state correlations among cooperating nodes that utilize shared positions and relative range information. Clearly, inter-nodal correlation may lead to non-converging positioning solutions if not accounted for. A Covariance Intersection Filter (CIF) approach attempts to mitigate the effect of unknown correlations by combining multiple estimates of state variables in the form of means and covariances, assuming that regardless of whether their correlation is unknown, the variables are always correlated [50,51]. The extension of the CIF concept to the Split-Covariance Intersection Filter (SCIF) aims to support this generalization of the correlation by splitting dependent (i.e., position and variance) and independent information (i.e., ranges and error) before the covariance intersection estimation [52]. As the SCIF may be implemented in the form of a modified KF, low complexity is ensured for multiple node position estimation, offering an attractive alternative for CP networks. A limitation of SCIF is identified in the requirement for the relative position to be known among cooperating nodes for successful solution convergence. Table 2 summarizes the main strengths and weaknesses for the EKF, PF, SPAWN, and CIS/SCIF algorithms. 4. Range Correction Models Section 4 presents the methodological framework for the development of range correction models. Based on the statistical measures obtained using UWB and Wi-Fi RTT observables, we propose distinct correction models and describe their respective implementation steps. Model validation procedures are established, and the associated developed software is presented. 4.1. Correction Models Development Steps The methodology followed for the design, development, implementation, and evaluation of the TWR range correction models relies on distinct steps as described in the following. These are developed as follows:Statistical characterization of range errors: At a first stage and before any range error modeling is applied, the raw TWR measurements undergo preliminary statistical analysis. For this purpose, a number of specifically designed experiments take place using an accurately surveyed testbed. In this regard, the statistics of the raw ranges carry useful information supporting the follow-up step of developing the data-driven range error models.Empirical range error models development: Prior to defining a range error model, the statistical metrics of the raw data obtained in the previous step are evaluated to assist in data grouping and data exclusion, or even dictate further data collection. The generation of error models resides on data-driven optimization techniques using regression analysis tools (best-fitting curves, interpolation, etc.). Statistical evaluation of the models before the implementation on real data enables the identification of potential gross deviations and data over-fitting.Error mitigation: This step includes implementation of the error models on real range data. Obviously, in order to obtain an unbiased evaluation, data correction refers to data collected only for validation purposes, excluding all data used for building the error model. By design, and in order for the error models to be efficient, they are classified to suit different operational conditions, usually by room type and geometry. Validation of the efficiency of error modeling is undertaken using a suitable subset of static reference data.Kinematic positioning: The final step of the range correction methodology concerns the error model performance assessment at the operational level. The dynamic character of the kinematic data provides the most demanding conditions for the TWR range error model validation due to the varying environment and user kinematics. For validating their reliability and robustness, a set of test trajectories is built that covers the entire testbed area. Statistical characterization of range errors: At a first stage and before any range error modeling is applied, the raw TWR measurements undergo preliminary statistical analysis. For this purpose, a number of specifically designed experiments take place using an accurately surveyed testbed. In this regard, the statistics of the raw ranges carry useful information supporting the follow-up step of developing the data-driven range error models. Statistical characterization of range errors: At a first stage and before any range error modeling is applied, the raw TWR measurements undergo preliminary statistical analysis. For this purpose, a number of specifically designed experiments take place using an accurately surveyed testbed. In this regard, the statistics of the raw ranges carry useful information supporting the follow-up step of developing the data-driven range error models. Empirical range error models development: Prior to defining a range error model, the statistical metrics of the raw data obtained in the previous step are evaluated to assist in data grouping and data exclusion, or even dictate further data collection. The generation of error models resides on data-driven optimization techniques using regression analysis tools (best-fitting curves, interpolation, etc.). Statistical evaluation of the models before the implementation on real data enables the identification of potential gross deviations and data over-fitting. Empirical range error models development: Prior to defining a range error model, the statistical metrics of the raw data obtained in the previous step are evaluated to assist in data grouping and data exclusion, or even dictate further data collection. The generation of error models resides on data-driven optimization techniques using regression analysis tools (best-fitting curves, interpolation, etc.). Statistical evaluation of the models before the implementation on real data enables the identification of potential gross deviations and data over-fitting. Error mitigation: This step includes implementation of the error models on real range data. Obviously, in order to obtain an unbiased evaluation, data correction refers to data collected only for validation purposes, excluding all data used for building the error model. By design, and in order for the error models to be efficient, they are classified to suit different operational conditions, usually by room type and geometry. Validation of the efficiency of error modeling is undertaken using a suitable subset of static reference data. Error mitigation: This step includes implementation of the error models on real range data. Obviously, in order to obtain an unbiased evaluation, data correction refers to data collected only for validation purposes, excluding all data used for building the error model. By design, and in order for the error models to be efficient, they are classified to suit different operational conditions, usually by room type and geometry. Validation of the efficiency of error modeling is undertaken using a suitable subset of static reference data. Kinematic positioning: The final step of the range correction methodology concerns the error model performance assessment at the operational level. The dynamic character of the kinematic data provides the most demanding conditions for the TWR range error model validation due to the varying environment and user kinematics. For validating their reliability and robustness, a set of test trajectories is built that covers the entire testbed area. Kinematic positioning: The final step of the range correction methodology concerns the error model performance assessment at the operational level. The dynamic character of the kinematic data provides the most demanding conditions for the TWR range error model validation due to the varying environment and user kinematics. For validating their reliability and robustness, a set of test trajectories is built that covers the entire testbed area. Figure 3 depicts the overall range correction procedure adopted in this study. Preliminary analyses suggest that TWR measurements do not necessarily follow a normal distribution indoors for reasons relating to multipath and through-material propagation effects. Therefore, the selection of a suitable statistical value is suggested. Figure 4 and Figure 5 show typical histograms of range datasets collected for the case of UWB and Wi-Fi RTT devices, respectively, for indoor environment conditions. Clearly, the histograms in the two figures indicate that the mean value cannot adequately represent the range sample. Moreover, while the median value provides a somehow improved index using the Empirical Probability Density Function (EPDF) for defining, the maximum likelihood value (EPDFmax) provides an optimal fit. The EPDFmax values need to be estimated, given that the respective histograms may not be utilized as a probability measure since they consist of discrete values (bins) that result in varying shapes based on the different bin sizes. The EPDF is estimated using kernel density estimation. It is crucial to select appropriate kernel bandwidth values, as larger bandwidth values smooth out the relevant peaks of EPDF, whereas for very small values, the remaining overall fluctuation hinders correct EPDFmax value estimation [23]. The empirically estimated kernel bandwidth value of 0.005 results in a good fit for the UWB data using the P410 module (Time Domain©, Huntsville, AL, USA), whilst the selection of a kernel bandwidth value of 0.02 results in a good fit for the WILD module (Compulab©, Yokne’am Illit, Israel) Wi-Fi RTT ranges. 4.2. Range Correction Models Following previous studies, the correction process for TWR data could be based either on empirical radial corrections, applying a least squares line fit to the range deviations as a function of the distance [23], or using a 2D range deviations plane fit [38]. In this study, we examine both approaches and extend the examination to WILD Wi-Fi RTT data in order to select the appropriate correction technique that suits the corresponding dataset. 4.2.1. Radial (1D) Fitting Model The development of a radial (1D) range correction model assumes the collection of TWR data at known (reference) distances using the RF devices of interest. For each pair of RF-ranging devices, a set of range measurements is collected to estimate their statistics and their deviation from the reference value. Hence, the correction value computes the difference between the one-way uncorrected measurement from the true (reference) distance as follows:(1)rangecorrection=rangetrue−rangemeasured Obviously, the ranges correction reflects the operational characteristics of the RF devices and the observation conditions applied in the area zone between the RF devices in use. The correction values may be estimated for various inter-device conditions in order to examine different environmental effects (i.e., NLOS conditions). The range correction models are realized through curve fitting on field data. Depending on the individual characteristics of the specific TWR technology and environmental conditions, different fit models may apply for each approach. The type correction models usually adopted are the “mean”, the “linear”, and the “polynomial” (second-order polynomial) fit. Figure 6 illustrates examples of various empirical correction models for UWB measurements. Notwithstanding the “polynomial” model appearing to more closely describe the nature of the range correction, a thorough examination is required in order to avoid over-fitting effects. Within this work, the models adopted refer to a linear fitting approach, as it has proved to better describe the collected TWR data, avoiding over-fitting effects. Figure 7 provides a schematic view of the procedure for empirical range correction models generation, outlining the distinct steps. For the case of a radial (1D) correction model, two variations are considered in this work, a generic linear correction model that covers all examined area, and a segmentation-based linear correction model to improve spatial resolution at a room level. The structure of the segmentation-based approach relies on the distribution of correction points in corresponding rooms. The first range correction approach (All Rooms Linear Correction, “arlc”) produces radial corrections for the complete test area irrespective of the room characteristics, and therefore, no distinction is made between LOS and NLOS conditions. The corresponding range correction equation reads (2)arlcin=din+fl(din) where fl is the linear range correction equation for all rooms for anchor node n. The second range correction variation (Room Linear Correction, “rlc”) produces a linear approximation of the correction values individually for each room depending on the continuously LoS or NLoS ranging conditions to specific anchor nodes each time. For instance, considering the case of Figure 7, the correction model for the left room corresponds solely to LOS ranging for anchors A1 and A2, and to NLOS ranging for anchors A3 and A4. The equation describing rlc reads (3)rlcin=din+fL(din,j) where din is the current (i) measured range between the roving node and anchor node n, and fL is the linear range correction equation for room j. 4.2.2. Spatial (2D) Fitting Model The generation of the two-dimensional range correction approach is based on the same underlying principle as the 1D approach (see Figure 8). In essence, the differences between the measured and true (reference) distances are used for the generation of a correction database connecting the correction points. In comparison to the linear fitting model, this approach takes into account the spatial distribution of the test ranges in the area of interest. Therefore, this method provides a bi-dimensional correction fit, which accounts for the location of each correction point. In order to cover the entire area of a test site, the correction values are interpolated using natural neighbor interpolation [53], which is based on the Voronoi tessellation method; hence, this Voronoi correction approach is denoted as “vc”. For the area found outside the polygons defined by the correction points, linear extrapolation is performed in order to extend the Voronoi correction values. The equation describing vc reads (4)vcin=din+fv(xin,yin) where fv is the bi-dimensional range corrections equation for the moving node’s position (xi,yi) for anchor node n. 4.3. Orientation-Assisted Range Correction Models 4.3.1. Orientation Assist Among the most influencing drawbacks concerned with TWR observables indoors are NLOS effects generated by physical obstacles or the multipath. In an attempt to initially model and consequently mitigate the effect of NLOS conditions in TWR ranges, orientation-assisted range error modeling is conceptualized and evaluated. In this regard, the most influencing factor associated with NLOS conditions for pedestrian indoor positioning is the same pedestrian’s body acting as a live obstacle. In order to examine and evaluate in a systematic manner the user orientation effect in relation to the anchor point, the data collection campaigns’ resolution described before in Section 4.2 is further increased by introducing the collection of discrete ranging datasets for all four cardinal orientations (north, east, south, and west). According to the rlc correction model, this approach generates a linear approximation of the correction values for each orientation. The orientation–linear correction model ("olc") is described by the equation (5)olcin=din+for(din,or) where din is the current (i) measured range between the roving node and anchor node n. Moreover, the expansion of the spatial (2D) correction model in order to include an additional level of detail based on the orientation assist is proposed and can be formulated as the orientation–Voronoi correction model (“ovc”) and is defined as follows:(6)ovcin=din+fov(xin,yin,or) where fov is the bi-dimensional range correction equation for the moving node’s n position (xi,yi) and for each orientation. 4.3.2. RSS-Based Orientation Selection In order to apply the correction models discussed in Section 4.3.1 in real case scenarios, the user orientation should be known. Notwithstanding today’s technology (e.g., MEMS IMU) being able to compute user orientation, at this stage, we exercise an autonomous RF-based approach. The proposed approach relies (1) on the provided data of each RF-based conversation, including both TWR observables along with signal quality information (RSS), and (2) on the hypothesis that the main source of RSS fluctuation for an otherwise static rover is the change in orientation due to the imposed NLOS conditions. Therefore, user orientation estimation relies on the comparison of the collected real-time RSS values against those obtained from previously collected RSS values for consequently selecting the appropriate orientation-based correction model. For this purpose, in addition to the linear and bi-dimensional models generated for the TWR measurables, the database is also populated with RSS-based linear and bi-dimensional models that are generated in a similar manner. For the case of the “olc” model, the RSS values are employed for generating a corresponding linear model for all anchor–rover pairs with respect to the reported range. These RSS models are then used during the online phase of the range correction algorithm by comparing the reported RSS value with respect to the reported uncorrected range and consequently selecting the closer RSS model. Similarly, for the case of the “ovc” model, the corresponding spatial RSS models are generated, and the real RSS values are compared against them in order to select the closer RSSI model and consequently the most, respectively, “ovc”-type model. Figure 9 illustrates the outline of the described RSS-based orientation selection approach. 4.4. Range Correction Models Validation In order to evaluate the appropriateness and operational efficiency of the range correction models, certain validation approaches are implemented. At the first stage, correction model validation refers to static ranges, aiming at computing detailed statistical measures whilst at the same time providing initial feedback for adopting a suitable correction model for the kinematic case. The second stage deals with the model validation process intended for kinematic positioning, specifically for evaluating range error mitigation effects under realistic positioning scenarios. 4.4.1. Internal and External Parameters Affecting TWR Quality Due to the inherent characteristics of the TWR observables and indoor environment conditions, which are of prime interest in this work, several factors need to be accounted for at the model validation stage. The internal factors effect refers to the varying setups that the TWR sensors may provide to the user, such as different signal transmission configuration values and sampling rate. The choice of the signal transmission configuration parameters such as the signal bandwidth or Pulse Integration Index (PII) affects ranging performance. Specifically, variations in signal configuration might provide the ability to acquire effectively range measurements over long distances and in return operate in lower sampling frequencies. Moreover, different recording bandwidths may provide variable ranging repeatability (i.e., precision) and multipath effects or NLOS resilience. Additionally, the choice of sampling rate values directly affects the positioning solution performance since, for example, a low sampling rate may hinder the ability to track motions of higher dynamics. On the other hand, a very high sampling rate might impede the localization engines of the roving nodes network, as it requires higher processing power in order to manage the increased data throughput. External effects refer to variations in the environmental conditions when performing TWR positioning. The indoor environment complex geometry, the presence of surrounding obstacles (static or mobile), and a user body acting as the main source of NLOS, are some of the determinant external factors. In addition, RF signal attenuation, scattering, and fading need to be accounted for and evaluated within a validation procedure. The different TWR technologies adopted in this research are expected to provide somewhat varying performance in varying environmental setups. Therefore, a detailed analysis takes place in order to gain insight that will facilitate subsequent experimental evaluation of positioning using a combination of the technologies. The NLOS being the main ranging quality degradation effect is examined using both through-the-wall TWR observables as well as the user’s body in a controlled and repeatable manner. 4.4.2. Validation Procedure of the Static Range Correction Model The validation of the static range correction model presupposes a series of suitable range datasets collected at different observation distances. This enables the statistical characterization of the raw observables, leading to conclusions about the performance of the correction models. The static validation datasets are collected in the same environment as the correction datasets since the ad hoc error correction models are suited for similar environmental conditions. Notably, performance assessment of the range correction models in variable environments exceeds the scope of this research. Naturally, in order to reach unbiased model assessment, the evaluation of the validation datasets is performed on data collected specifically for validation purposes and not on those collected for error modeling. The number of validation points selected ranges between 30 and 40% of the total datasets points, which is adequate for providing reliable evaluation results. The field procedure includes range observation in a 1D setup from an anchor point to a rover placed sequentially at increasing distances along a corridor-like geometry. Also, the procedure may be performed in a 2D setup. In this scenario, the observation points are spatially distributed throughout the area of interest, and the corresponding Euclidean distances are computed based on the known anchor coordinates. At the implementation stage, the radial and spatial correction models and associated software are implemented. Subsequently, the corrected ranges are cross-compared against the nominal distances resulting in a statistical evaluation (i.e., trueness, mean, and standard deviation) for gaining insight regarding the parameters analyzed in Section 4.4.1. Corresponding trueness histograms facilitate the quantitative performance evaluation for each pair. Based on the corrected ranges, the remaining error EPDFmax value may be used for all validation points in order to generate remaining error diagrams in contour or heatmap form. 4.4.3. Validation Procedure of the Kinematic Range Correction Model Since the aim is to enable a correction model for kinematic (dynamic) range evaluation for real-time applications, the validation procedure needs to be expanded for the kinematic case. This validation mode intends to evaluate the performance of the developed models in a realistic manner, whilst at the same time coping with fewer observables than the static one per rover position, and therefore no detailed statistical measures can be estimated. Usually, the estimation of a reference trajectory indoors relies on the realization of a predefined path along previously established and accurately surveyed points. Positioning performance evaluation relies on the comparison of the estimated trajectories performance using the different correction models. Moreover, the assumptions underlying each model implementation are different, considering that the radial (1D) models rely only on the measured range, while the spatial (2D) models rely on the previously estimated position. This validation step allows for the evaluation of the model implementation in real TWR datasets intended for trajectory estimation. The trajectory quality metrics estimated against the reference trajectory enable the quantitative comparison among varying models. 4.5. Developed TWR Correction and Validation Software Within the scope of designing and implementing a generalized approach for building TWR range correction models, a dedicated software (SW) suite has been designed and developed in Matlab® R2020a Programming Environment. The SW suite receives as input suitably formatted raw TWR data; it generates the corresponding correction models, performs validation checks, and finally generates validation statistics tables [8]. 5. Position Computation Algorithm The goal underlying this study is the development and evaluation of a suite of decentralized CP algorithms to enable the localization of multiple rovers using RF-based TWR observables collected in a network of roving and static nodes architecture. The basis of the absolute localization engine relies on Extended Kalman Filtering (EKF) realized in a collaborative manner. Considering that the adoption of a CP strategy entails the introduction of uncertainty due to the correlated positioning solutions, it is expected to affect the network solution, resulting in highly inaccurate results or even an inability for filter convergence. In an attempt to optimally combine Pedestrian-to-Pedestrian (P2P) range measurements in a decentralized manner, an approach is formulated based on Split Covariance Intersection (SCI) grounds using the inter-device TWR ranges, the advertised rover state, and covariance information. For the localization of a mobile rover using P2I ranges, the observation setup relies on the provision of TWR observables from anchors of known coordinates to the rover in a dynamic manner. The range measurements are processed sequentially upon recording along with the reported accuracy (as estimated by the device) and the system timestamp. In a scenario of multiple rovers, each rover utilizes independently its corresponding measurements as they become available. Figure 10 illustrates the basic system setup for a single roving pedestrian and four ranges captured sequentially from the four anchors. Notwithstanding this experimental setup referring to a single rover, it may support multiple rovers subject to a potential limitation imposed by the maximum TWR technology communication network capacity. The key elements of the EKF are the state vector (X), covariance (P), system model or state transition function (f), process noise (Q), measurement vector (Z), measurement covariance (R), measurement model (H), and the Kalman gain (K). Following typical approaches, the EKF is implemented as follows: Initialization (7) X^0=[X0] (8) P0≅Q Initialization (7) X^0=[X0] (8) P0≅Q Prediction (9) X^t¯=FX^t−1 (10) Pt¯=FPt−1FT+Q Prediction (9) X^t¯=FX^t−1 (10) Pt¯=FPt−1FT+Q Correction (11) Kt=Pt¯HT(HPt¯HT+R)−1 (12) X^t=X^t¯+Kt(Zt−HX^t¯) (13) Pt=(I−KtH)Pt¯ Correction (11) Kt=Pt¯HT(HPt¯HT+R)−1 (12) X^t=X^t¯+Kt(Zt−HX^t¯) (13) Pt=(I−KtH)Pt¯ 5.1. Correction Models Adopted for the UWB and Wi-Fi RTT Range Observables Internal Accuracy The TWR technologies employed in this study (i.e., Time Domain© P410 UWB, Compulab© WILD Wi-Fi RTT) result in range measurements with device-generated error values. In this study, two schemes for providing Dynamic Measurement Error Estimation (DME) algorithms are adopted and presented. Figure 11 depicts in red the range differences obtained between the observed and reference values for a UWB rover module using a preliminary data campaign featuring four UWB anchors. Also, the same plots show in blue the median values (Leading-Edge Detection, LED) and their associated standard deviations as recorded by the sensors. According to the manufacturer, the LED flag value for LoS conditions should equal eight, whilst larger values indicate NLOS operation. Moreover, from the same plots, a relationship between the recorded LED values and the corresponding range deviations is observed, indicating that the reported values can be utilized as an index for characterizing the range quality together with their reported range error values. In this regard, the noise of range measurements is defined by the error value reported by the UWB module for each measurement. Thorough examination of the reported (by the sensor) range errors against their estimated equivalents (computed standard deviation) suggests the introduction of a scaling factor to the reported error. Based on the relationship between the reported LED values against the computed range accuracy, an empirical scaling tactic is engaged during real-time ranging as described by the following equation:(14)σtrU=errtr∗5,7