[Paper Review] Distributed Fusion of Labeled Multi-Object Densities Via Label Spaces Matching

Bailu Wang, Wei Yi+3 more|arXiv (Cornell University)|Mar 28, 2016

Target Tracking and Data Fusion in Sensor Networks12 references3 citations

TL;DR

This paper proposes GCI-LSM, a two-step distributed fusion algorithm for labeled multi-object densities that addresses label space mismatching (LS-DM) by first matching label spaces using a ranked assignment problem based on an information divergence criterion (LS-M), then applying Generalized Covariance Intersection (GCI) fusion. The method significantly improves tracking performance over direct fusion when sensors use inconsistent labeling, especially in scenarios with target births and deaths.

ABSTRACT

In this paper, we address the problem of the distributed multi-target tracking with labeled set filters in the framework of Generalized Covariance Intersection (GCI). Our analyses show that the label space mismatching (LS-DM) phenomenon, which means the same realization drawn from label spaces of different sensors does not have the same implication, is quite common in practical scenarios and may bring serious problems. Our contributions are two-fold. Firstly, we provide a principled mathematical definition of "label spaces matching (LS-DM)" based on information divergence, which is also referred to as LS-M criterion. Then, to handle the LS-DM, we propose a novel two-step distributed fusion algorithm, named as GCI fusion via label spaces matching (GCI-LSM). The first step is to match the label spaces from different sensors. To this end, we build a ranked assignment problem and design a cost function consistent with LS-M criterion to seek the optimal solution of matching correspondence between label spaces of different sensors. The second step is to perform the GCI fusion on the matched label space. We also derive the GCI fusion with generic labeled multi-object (LMO) densities based on LS-M, which is the foundation of labeled distributed fusion algorithms. Simulation results for Gaussian mixture implementation highlight the performance of the proposed GCI-LSM algorithm in two different tracking scenarios.

Motivation & Objective

• To address the critical problem of label space mismatching (LS-DM) in distributed multi-target tracking, where identical realizations from different sensors' label spaces have inconsistent meanings.
• To provide a principled mathematical definition of label space matching (LS-M) based on information divergence, enabling a formal criterion for matching quality.
• To develop a robust two-step fusion framework that first aligns label spaces across sensors and then performs GCI fusion on the matched space.
• To derive a general GCI fusion rule for labeled multi-object (LMO) densities under the LS-M criterion, enabling broader applicability to various LMO filters.
• To demonstrate the superiority of the proposed method over direct fusion with mismatched labels, particularly in scenarios with target births, deaths, and dynamic labeling.

Proposed method

• Proposes a principled LS-M criterion based on information divergence to define when two label spaces are considered matched.
• Models the label space matching as a ranked assignment problem, where the cost function is derived from the LS-M criterion to find optimal label correspondences between sensors.
• Uses a Gaussian mixture implementation to efficiently compute the LS-M cost function and solve the assignment problem in practice.
• Applies Generalized Covariance Intersection (GCI) fusion on the matched label space, ensuring robust fusion even with unknown cross-correlations.
• Derives a general GCI fusion rule for generic LMO densities under the LS-M assumption, forming the theoretical foundation for the algorithm.
• Employs adaptive birth models and prior information on object births to improve performance in scenarios with dynamic target appearance.

Experimental results

Research questions

• RQ1 How can label space mismatching (LS-DM) be formally defined and quantified in distributed multi-target tracking?
• RQ2 What is a principled, information-theoretic criterion for determining whether two label spaces are matched?
• RQ3 How can optimal label correspondence between sensors be found when label spaces are mismatched?
• RQ4 Can GCI fusion be effectively applied to labeled multi-object densities after resolving LS-DM?
• RQ5 How does the proposed GCI-LSM algorithm perform compared to direct fusion methods in scenarios with target births and deaths?

Key findings

• The GCI-LSM algorithm significantly outperforms direct GCI-LMB fusion in scenarios with target births and deaths, where label space mismatching causes complete loss of tracks.
• In the PBP scenario, GCI-LSM maintains accurate tracking of all targets, while GCI-LMB fusion fails to fuse the second track due to label mismatching.
• In the five-target scenario, GCI-LSM achieves OSPA performance nearly identical to GCI-GMB fusion after convergence, demonstrating near-optimality of the label matching step.
• The GCI-LSM fusion performs slightly worse than GCI-GMB fusion during target birth phases, but this is attributed to the use of a single optimal assignment rather than full joint matching, confirming the accuracy of the ranked assignment model.
• Both GCI-LSM and GCI-GMB fusion outperform GCI-PHD fusion in both OSPA error and cardinality estimation, highlighting the advantage of using labeled densities.
• The simulation results confirm that resolving LS-DM is essential for robust distributed multi-target tracking, especially in dynamic environments with frequent target spawning and termination.