Probability hypothesis density filter using determinantal point processes for multi object tracking

摘要

Multi Object Tracking (MOT) has many applications such as video surveillance and event recognition among others. In this paper, we present a novel multi object tracking method using the Probability Density Hypothesis (PHD) filter and Determinantal Point Processes (DPP). The PHD filter is an algorithm for jointly estimating an unknown number of targets and their states from a sequence of observations in the presence of data association uncertainty, noise and false alarms. A tractable implementation of the PHD filter is based on a Gaussian Mixture approximation. However, the Gaussian Mixture PHD suffers from computational problems due to an increasing number of Gaussian components as time progresses. In this paper, we propose a novel pruning method based on Determinantal Point Process which handles the overestimation problem on the number of tracks. The DPP-PHD filter promotes diversity in the resulting Gaussian components and leads to improved tracking results.