[Paper Review] Distributed Computation Particle PHD filter

Junjie Wang, Lingling Zhao+2 more|arXiv (Cornell University)|Mar 9, 2015

Target Tracking and Data Fusion in Sensor Networks3 references3 citations

TL;DR

This paper proposes a distributed computation particle PHD (DCPPHD) filter that accelerates multi-target tracking by partitioning particles across multiple processing elements (PEs) while maintaining accuracy. Each PE runs a local particle PHD filter using full measurements and exchanges particles with a central unit, which fuses estimates via a fusion rule, achieving significant speedup over centralized particle PHD filtering with comparable tracking performance.

ABSTRACT

Particle probability hypothesis density filtering has become a promising means for multi-target tracking due to its capability of handling an unknown and time-varying number of targets in non-linear non-Gaussian system. However, its computational complexity grows linearly with the number of measurements and particles assigned to each target, and this can be very time consuming especially when numerous targets and clutter exist in the surveillance region. Addressing this issue, we present a distributed computation particle PHD filter for target tracking. Its framework consists of several local particle PHD filters at each processing element and a central unit. Each processing element takes responsibility for part particles but full measurements and provides local estimates; central unit controls particle exchange between processing elements and specifies a fusion rule to match and fuse the estimates from different local filters. The proposed framework is suitable for parallel implementation and maintains the tracking accuracy. Simulations verify the proposed method can provide comparative accuracy as well as a significant speedup with the standard particle PHD filter.

Motivation & Objective

• To address the high computational complexity of centralized particle PHD filters in nonlinear, non-Gaussian, and cluttered environments.
• To enable real-time multi-target tracking by distributing particle computation across multiple processing elements.
• To maintain tracking accuracy comparable to the standard particle PHD filter while improving runtime efficiency.
• To develop a scalable framework suitable for parallel and distributed implementation in sensor networks or high-performance computing.

Proposed method

• The framework distributes N particles across P processing elements (PEs), each running an independent local particle PHD filter using all available measurements.
• Each PE computes local estimates based on its assigned particles and transmits them to a central unit (CU).
• The central unit manages inter-PE particle exchange and applies a fusion rule to match and combine local estimates into a global estimate.
• The STPHD method is used to extract state estimates and measurement labels, improving estimation accuracy.
• A particle exchange mechanism prevents performance degradation due to limited particle sets per PE.
• The method leverages SMC (Sequential Monte Carlo) sampling and PHD recursion for prediction and update steps.

Experimental results

Research questions

• RQ1 Can a distributed architecture reduce the computational load of particle PHD filtering without degrading tracking accuracy?
• RQ2 How does particle distribution across multiple PEs affect estimation performance and convergence in multi-target tracking?
• RQ3 What is the impact of inter-PE particle exchange on maintaining estimation accuracy in a distributed setup?
• RQ4 Can the proposed method achieve significant speedup compared to centralized particle PHD filtering under equivalent particle counts?

Key findings

• The DCPPHD filter achieved a 50% reduction in average execution time (8.63s vs. 17.65s) compared to the centralized particle PHD filter when using the same total number of particles (2000).
• The mean OSPA distance for DCPPHD was 3.0732, closely matching the centralized particle PHD filter (3.0577), indicating comparable tracking accuracy.
• With 1/5 the number of particles per PE (500 vs. 400), the DCPPHD achieved similar OSPA performance to a centralized filter with only 1/40,000 particles per PE, demonstrating efficiency.
• The method maintained consistent target number estimation, with estimated target counts closely tracking true trajectories across 50 scans.
• The framework showed scalability, with performance degradation observed when increasing the number of processing groups beyond optimal levels.