Data Association in Multiple Object Tracking: Challenges and Solutions

Abstract: Multi-object tracking plays a crucial role in a variety of fields, including autonomous systems, defense, and biology. A central challenge in multi-object tracking is the data association problem, arising from the unknown correspondence between objects and noisy sensor measurements. In this talk, we will first introduce the multi-object tracking problem and its model-based Bayesian formulation using random finite sets of trajectories. We will then present the multi-object posterior in the form of a Poisson multi-Bernoulli mixture (PMBM) and explain how data association hypotheses are captured within this filtering framework. Following this, we will describe two popular PMBM algorithms—one hypothesis-oriented and one track-oriented—highlighting how the data association problem can be formulated as an optimization problem and efficiently solved. Finally, we will discuss performance evaluation for multi-object tracking, including the associated optimisation challenges.