Robust Object Tracking with Continuous Data Association based on Artificial Potential Moving Horizon Estimation

Abstract

Abstract In this paper, a novel object tracking method based on moving horizon estimation (MHE) is introduced that integrates data association into numerical optimization on Bayesian state estimation. Object tracking is a classical problem that often appears in radar and vision systems for which either deterministic or probabilistic approach has been applied. While the former often encounter a failure of association, the latter avoids it by computing the expectation concerning the observations, but it requires the prior knowledge of the probabilistic distribution and may suffer from outliers. In this paper, a partially deterministic approach is built on MHE to resolve these concerns. Data association is realized by a potential function comprising the superposition of Gaussian distributions centered at each observed feature. The potential function is embedded into the stage cost of MHE; the optimal data association is conducted using the sequence of observations within the horizon. Thus robust object tracking is achieved utilizing multi-sampling data and integration of both dynamics and explicit constraints reflecting physical limitations. The advantage of the proposed method is verified through object tracking experiments on the crowded environment, where occlusion and misrecognition frequently occur.