Multi-target Tracking and Segmentation via Discriminative Appearance Model

we present a novel discriminative appearance model for monocular multi-target tracking and segmentation in a comparatively crowded scene. Based on the hypothesis that the discriminability among different targets plays an important role in improving the tracking performance, we choose different feature spaces for every target in the scene to insure the discriminability from other targets. In order to adapt to continuously changing appearance, we propose to adjust the updating ratio of the model according to the change of motion direction. We propose a two-level tracking algorithm to track and segment multi-target, which integrates our discriminative appearance model into a probabilistic data association framework. Our tracking algorithm is more effective and efficient. Tracking results on the public dataset PETS2009, compared with the conventional appearance model in the same feature space, show a great improvement, especially in segmenting much more accurately during occlusions and reducing identity switches more significantly.