Multiple Target Tracking in Unknown and Time-Varying Measurement Noise Intensity

Tracking multi-object states under time-varying (TV) measurement noise presents a critical challenge in numerous real-world applications, from radar tracking to sensor networks, where dynamic environments and uncertain observations are the norm. Accurately capturing the interplay between object states and evolving noise parameters is essential for robust performance, yet existing methods often fall short when handling unknown and time-dependent noise variations. In this paper, we address this gap by introducing a novel framework based on a dependent Dirichlet process (DDP), which flexibly models noise parameters as marginally distributed at each time step and jointly estimates states and noise dynamics. The approach not only accommodates an unknown and varying number of objects but also detects temporal shifts in noise characteristics. We employ a Gibbs sampling-based strategy for efficient posterior computation; the proposed method achieves reliable and adaptive performance in complex, non-stationary environments. We demonstrate through experiments the robustness and performance of our proposed methods.