Tracking by Risky Particle Filtering over Sensor Networks

Jaechan Lim, Hyung‐Min ParkSensors 2020 · DOI

TL;DR

This paper introduces "minimax particle filtering" (MPF) to improve target tracking in wireless sensor networks using Received Signal Strength (RSS) measurements. By adopting a "maximum risk" approach when calculating particle weights, MPF significantly reduces weight variance, mitigating the common "degeneracy problem" in standard particle filters and leading to more robust and accurate tracking.

What Problem Does This Paper Address?

Wireless sensor networks (WSNs) are crucial for applications like the Internet of Things (IoT), enabling smart systems that collect and transmit data without direct human intervention. A key challenge in WSNs is accurately tracking moving targets, especially when relying on Received Signal Strength (RSS) measurements from multiple sensors. Standard particle filtering (PF) methods, while powerful, suffer from a "degeneracy phenomenon" where particle weights become highly uneven over time, leading to poor tracking performance. This issue is exacerbated in scenarios with high target dynamics or limited sensor data. The paper aims to address this by developing a more robust particle filtering approach that maintains particle diversity and improves tracking accuracy.

Methods

The core innovation is the minimax particle filtering (MPF) approach. Unlike standard particle filters that aim to minimize the expected squared error (MMSE), MPF seeks to minimize the maximum risk. This is achieved by modifying the particle weight calculation. Instead of using all available RSS measurements to compute a single weight, MPF strategically selects one measurement that, when used to calculate the weight, results in the minimum weight for that particle. This counter-intuitive approach, by devaluing individual particle contributions based on a single, potentially "risky" measurement, effectively reduces the overall variance of particle weights. This strategy is then applied to various established PF variants, including Standard PF (SPF), Auxiliary PF (APF), Regularized PF (RPF), Kullback-Leibler divergence PF (KLDPF), and Gaussian PF (GPF), creating their respective minimax counterparts (e.g., MPF, MAPF, MRPF, MKLDPF, MGPF).

Datasets

Not discussed in the paper. The experiments were conducted using simulated data generated based on the defined dynamic state and measurement models.

Key Insights and Ideas

  • Minimax Risk for Weighting: The core idea of minimizing the maximum risk, rather than the expected risk, is a novel way to stabilize particle filter weights. It's a counter-intuitive approach that prioritizes robustness over immediate optimal performance for each particle.
  • Weight Variance Reduction: The minimax strategy directly tackles the degeneracy problem by significantly reducing the variance of particle weights. This leads to better particle diversity and more stable tracking.
  • Scalability with Sensors: The benefits of MPF become more pronounced as the number of sensors (M) increases. With more sensors, the ability to strategically select a "risky" measurement to minimize a particle's weight becomes more effective.
  • Robustness in Dynamic Scenarios: MPF shows particular advantage in scenarios with higher state noise variance (S_L), where target dynamics are more severe and degeneracy is a greater concern.
  • Generalizability: The minimax strategy is shown to be effective across various standard PF implementations, demonstrating its broad applicability.

Key Results

  • Improved Tracking Accuracy: Minimax PF variants consistently outperformed their standard counterparts in terms of Mean Distance Error (MDE) and Mean Square Error of Distance (MSED), especially when using a larger number of sensors (M=5 and M=10).
  • Reduced Weight Variance: Figures across the paper (e.g., Figure 3, 6, 9, 12, 15) demonstrate that MPF significantly reduces the variance of particle weights compared to standard PFs. This reduction is more substantial as M increases.
  • Lower Mean Risk: The minimax approach generally resulted in lower mean-risk values, particularly with an increased number of sensors.
  • Performance Gains: While standard PFs showed similar performance with only two sensors, MPF variants showed clear superiority as M increased. The performance gap was more evident in scenarios with higher state noise (S_L).
  • Processing Time: The minimax strategy did not introduce significant additional computational overhead compared to standard PF variants, with some variants (like GPF) being inherently faster due to the absence of resampling.

Limitations and Future Work

  • Signal-to-Noise Ratio (SNR) Dependency: The paper notes that the minimax strategy is effective when the signal-to-noise ratio is above a certain threshold. Below this threshold, MPFs might not outperform regular PFs.
  • Computational Complexity of KLDPF: While the minimax strategy itself doesn't add significant overhead, the KLDPF variant it's applied to inherently has higher computational costs.
  • Future Work: The paper suggests that the proposed minimax strategy is highly effective and can be applied to other variants of PF. Further verification in more complex scenarios or with real-world data could be a natural next step. The authors also mention that performance improvement may saturate with a very large number of sensors due to computational complexity.

Summary updated

Abstract

The system of wireless sensor networks is high of interest due to a large number of demanded applications, such as the Internet of Things (IoT). The positioning of targets is one of crucial problems in wireless sensor networks. Particularly, in this paper, we propose minimax particle filtering (PF) for tracking a target in wireless sensor networks where multiple-RSS-measurements of received signal strength (RSS) are available at networked-sensors. The minimax PF adopts the maximum risk when computing the weights of particles, which results in the decreased variance of the weights and the immunity against the degeneracy problem of generic PF. Via the proposed approach, we can obtain improved tracking performance beyond the asymptotic-optimal performance of PF from a probabilistic perspective. We show the validity of the employed strategy in the applications of various PF variants, such as auxiliary-PF (APF), regularized-PF (RPF), Kullback-Leibler divergence-PF (KLDPF), and Gaussian-PF (GPF), besides the standard PF (SPF) in the problem of tracking a target in wireless sensor networks.
BibTeX
@inproceedings{lim2020tracking,
  author = {Jaechan Lim and Hyung‐Min Park},
  title = {{Tracking by Risky Particle Filtering over Sensor Networks}},
  booktitle = {Sensors},
  year = {2020},
  doi = {10.3390/s20113109},
}
Figure