On Pooling-Based Track Fusion Strategies : Harmonic Mean Density

Nikhil Sharma, Ratnasingham Tharmarasa, Shovan Bhaumik, Thiagalingam Kirubarajan · 2024

Read On Pooling-Based Track Fusion Strategies : Harmonic Mean Density by Nikhil Sharma, Ratnasingham Tharmarasa, Shovan Bhaumik, Thiagalingam Kirubarajan (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06716.txt.

Introduction

Nikhil Sharma, Shovan Bhaumik, Ratnasingham Tharmarasa, and Thiagalingam Kirubarajan fuse sensor tracks without falling into the traps of naive fusion. In distributed systems, plain filter fusion ignores how tracks correlate and degrades results, so cautious methods step in. Mixtures of Gaussians model densities flexibly, but the standard cautious method for them is cumbersome. The team builds a pooling strategy on the harmonic mean of local densities.

The same framework handles plain and mixture densities with little change. The authors study the method's mathematical properties and test it on tracking of maneuvering targets in two and three dimensions. It beat rival cautious strategies on error while staying consistent.

Catalog record

CREATORNikhil Sharma, Ratnasingham Tharmarasa, Shovan Bhaumik, Thiagalingam Kirubarajan
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE70 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06716

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

On Pooling-Based Track Fusion Strategies : Harmonic Mean Density Nikhil Sharma, Ratnasingham Tharmarasa, Shovan Bhaumik, Thiagalingam Kirubarajan Abstract In a distributed sensor fusion architecture, using standard Kalman filter (naive fusion) can lead to degraded results as track correlations are ignored and conservative fusion strategies are employed as a sub-optimal alternative to the problem. Since, Gaussian mixtures provide a flexible means of modeling any density, therefore fusion strategies suitable for use with Gaussian mixtures are needed. While the generalized covariance intersection (CI) provides a means to fuse Gaussian mixtures, the procedure is cumbersome and requires evaluating a non-integer power of the mixture density. In this paper, we develop a pooling-based fusion strategy using the harmonic mean density (HMD) interpolation of local densities and show that the proposed method can handle both Gaussian and mixture densities without much changes to the framework. Mathematical properties of the proposed fusion strategy are studied and simulated on 2D and 3D maneuvering target tracking scenarios. The simulations suggest that the proposed HMD fusion performs better than other conservative strategies in terms of root-mean-squared error while being consistent. Journal: Signal Processing † † graphicalabstract: Track to Track Fusion Exact Distributed Track Filtering [ 1 ] Channel Filter [ 2 ] Globalized Likelihood [ 3 ] Decorrelation Tracklets [ 4 ] Quasi-Tracklets [ 5 ] IMM Decorrelation [ 6 ] Pooling Strategies [ 7 ] Geometric Averaging [ 8 ] Covariance Intersection [ 9 ] Arithmetic Averaging [ 10 ] Harmonic Averaging Information form Generalized IMF [ 11 ] Sampling based Ensemble KF [ 12 ] 1 Introduction A distributed data fusion scenario comprises several sensor platforms connected to a node where the local information from sensors is fused. Since a sensor can capture multiple detections in the presence of clutter, sending raw measurement is avoided. Instead, the sensor communicates estimated target states (tracks), which is why the term track-to-track fusion (T2TF) is commonly adopted. Tracks estimating the same target trajectory in T2TF are never independent, even though the local measurement errors can be assumed uncorrelated [ 13 ] . The dependence emerges due to the common process noise, and temporal correlation between tracks arriving from the same platform. Ignoring this correlation can lead to a degraded performance as the resulting estimate is optimistic [ 14 ] . The optimism is due to the aggregation of “common information" in sensor estimates. This common information has to be accounted for and removed from the fused information to avoid track divergence. The approaches used in accounting for cross-correlation for the last three decades can be categorized into: 1. Optimal solution based on evaluating exact recursion for correlation, which can later be accounted for in fusion. This strategy would require extra information from local sensors like Kalman gain, sensor, and motion model parameters [ 13 ] [ 15 ] . 2. Conditionally optimal methods like information matrix fusion, which would also require knowledge of previous local estimates [ 16 ] . Such methods are optimal only for a full communication rate, i.e. the fusion occurs every time the local estimates are updated. 3. Tracklet and quasi-tracklet fusion methods are based on track decorrelation by marginalizing the joint density. Here again, knowledge about a track’s previous estimate is needed. 4. Pooling based solutions like the generalized covariance intersection, which utilizes a geometric mean (GM) or the arithmetic mean (AM) interpolation of individual track densities [ 17 , 10 ] . Generalized covariance intersection/Chernoff fusion/geometric averaging has attracted researchers due to its robustness and stability [ 18 ] . An essential aspect of this method is that it essentially constitutes a density function which is a log-linear combination of the individual local posterior densities [ 17 ] , such that the propagation of the “common information" is avoided [ 19 ] . Due to computational intensiveness of geometric mean for Gaussian mixtures, an alternative was proposed in [ 10 , 20 ] . With the advent of computation and embedded technologies, the use of mixture densities has been adopted to deal with multi-modal, non-symmetric noise distributions and multiple-model approaches to filtering. Moreover, the Gaussian mixture models (GMM) possess several standard results available in closed form. Some examples utilizing this advantage include the interacting multiple model (IMM) filter and its modifications for tracking maneuvering targets [ 21 ] , the multiple hypothesis tracker (MHT) for multiple target tracking [ 22 ] , Gaussian sum filter [ 23 ] and the probability hypothesis density (PHD) filter [ 24 ] . These filters are extensively implemented owing to their performance in realistic situations. However, the fusion of such cross-correlated mixture-modelled tracks becomes infeasible due to the unavailability of closed-form expressions. For instance, decorrelation requires division of prior and posterior densities, for which an exact result is impossible if Gaussian mixtures are involved [ 25 ] . Chernoff fusion requires computing non-integer powers of a Gaussian mixture which does not assume a closed-form [ 26 ] . The arithmetic averaging, though feasible, presents a relatively larger uncertainty due to amalgamation of densities. Therefore, further processing in the form of mode merging and pruning is needed [ 10 ] . Due to such reasons, fusion of Gaussian mixtures is non-trivial in practical scenarios, and we need to search in alternative directions. This article proposes a method which can handle uni-modal and multi-modal Gaussian densities without altering the framework. Research on Gaussian mixture fusion is contemporary and various solutions are available in literature. [ 27 ] addressed the issue of deriving the dependency relation between Gaussian mixture estimates and found that the dependency is itself a Gaussian mixture. No remarks on how to ease the fusion, however, were made. A tracklet [ 5 ] based fusion strategy for multiple model tracker is proposed in [ 6 ] where the information decorrelation approach was extended to Gaussian mixture densities and then naive fusion was applied on the decorrelated mixture. This work required division by a Gaussian mixture wherein, a crude approximation of component-wise division was applied. In [ 28 ] , the dynamic state of an IMM running fusion center was derived when the local tracks are in turn IMMs using the “inside information". But, this approach was based on the assumption that individual models in the IMM are entirely the same except for the process noise. Thus, the approach is not useful in scenarios where, for e.g. nearly-constant velocity (NCV) and nearly-constant acceleration (NCA) models are employed in parallel. Recently, conservative fusion methodology has taken a different direction where the use of Frétchet-means [ 29 ] are explored. These works are based on the fact that averaging rules are immune to double counting, thus imparting robustness with respect to correlations involved without requiring any extra information from the local trackers. In [ 7 ] , a formal definition of conservativeness was presented and proved for arithmetic and geometric averaging. The definition was based on the amount of uncertainty in the distribution, thus using measures such as the entropy and the Kullback-Leibler divergence. In [ 10 ] the arithmetic averaging was applied on a multitarget tracking scenario using the multi-Bernoulli process in which each component was represented by a Gaussian mixture. In [ 19 ] , the notion conservativeness was

Read the complete text