[Paper Review] Arithmetic Average Density Fusion -- Part II: Unified Derivation for Unlabeled and Labeled RFS Fusion

Tiancheng Li|arXiv (Cornell University)|Sep 21, 2022

Remote-Sensing Image Classification4 citations

TL;DR

This paper presents a theoretically unified and exact derivation of arithmetic average (AA) fusion for both unlabeled and labeled random finite set (RFS) filters in multi-sensor multi-target tracking. By introducing probability hypothesis density (PHD) consistency as a unifying principle, it enables robust, approximation-free fusion that enhances target detection and localization accuracy across heterogeneous RFS filters, including GLMB, LMB, and MB, without requiring full posterior matching or complex history coordination.

ABSTRACT

As a fundamental information fusion approach, the arithmetic average (AA) fusion has recently been investigated for various random finite set (RFS) filter fusion in the context of multi-sensor multi-target tracking. It is not a straightforward extension of the ordinary density-AA fusion to the RFS distribution but has to preserve the form of the fusing multi-target density. In this work, we first propose a statistical concept, probability hypothesis density (PHD) consistency, and explain how it can be achieved by the PHD-AA fusion and lead to more accurate and robust detection and localization of the present targets. This forms a both theoretically sound and technically meaningful reason for performing inter-filter PHD AA-fusion/consensus, while preserving the form of the fusing RFS filter. Then, we derive and analyze the proper AA fusion formulations for most existing unlabeled/labeled RFS filters basing on the (labeled) PHD-AA/consistency. These derivations are theoretically unified, exact, need no approximation and greatly enable heterogenous unlabeled and labeled RFS density fusion which is separately demonstrated in two consequent companion papers.

Motivation & Objective

• To address the lack of a theoretically unified framework for arithmetic average (AA) fusion across diverse unlabeled and labeled RFS filters in multi-sensor multi-target tracking.
• To establish PHD consistency as a principled foundation for AA fusion, explaining its robustness to false and missing detections.
• To derive exact, approximation-free AA fusion formulas for major RFS filters, including GLMB, LMB, and MB, without relying on MPD-best-fit heuristics.
• To reduce fusion complexity by eliminating the need for full history or measurement-track association matching across sensors.
• To enable practical, computationally efficient fusion of heterogeneous RFS filters through label-wise and existence-probability averaging.

Proposed method

• Introduces the concept of PHD consistency as a theoretical basis for AA fusion, ensuring fusion preserves the form of the fusing RFS filter and improves detection accuracy.
• Derives exact AA fusion formulas for key RFS filters (Bernoulli, PHD, CPHD, MB, MBM, GLMB, LMB, M-GLMB) by averaging their respective sufficient statistics (e.g., existence probabilities, spatial density functions).
• Applies label-wise fusion for labeled RFS filters (e.g., LMB, GLMB), requiring only label matching rather than full history or association hypothesis alignment.
• Uses the L2 norm metric for fusion in the labeled case, avoiding reliance on KL divergence or MPD-best-fit approximations.
• Employs coordinate descent and variational approximation techniques in companion works to enable efficient implementation of fused filters.
• Provides a unified framework where all RFS filter types are fused via averaging their (labeled) PHD components, ensuring consistency and computational scalability.

Experimental results

Research questions

• RQ1 How can arithmetic average fusion be rigorously justified for both unlabeled and labeled RFS filters in multi-sensor multi-target tracking?
• RQ2 What theoretical principle underlies the robustness of AA fusion to false and missing detections in multi-target scenarios?
• RQ3 Why do existing MPD-best-fit derivations fail to explain the performance of AA fusion in labeled RFS settings?
• RQ4 Can a unified, exact fusion framework be derived for heterogeneous RFS filters without relying on approximations or full posterior matching?
• RQ5 What are the minimal requirements (e.g., label matching vs. history matching) for successful fusion of labeled RFS filters?

Key findings

• The paper derives the first exact AA fusion formula for the general GLMB filter, resolving a long-standing gap in the literature.
• PHD consistency is established as a unifying principle that explains why AA fusion improves detection and localization accuracy without double-counting information.
• Labeled PHD fusion is made feasible with only label matching, eliminating the need for complex history or association hypothesis coordination across sensors.
• The method avoids MPD-best-fit approximations, which are theoretically flawed and fail to explain AA fusion’s empirical success.
• The framework enables efficient fusion of heterogeneous filters (e.g., PHD, MB, LMB) by averaging only their respective sufficient statistics, such as existence probabilities and spatial density functions.
• Open-access MATLAB codes are provided for AA-fusion-based RFS filters, supporting practical deployment and integration of heterogeneous sensor data.