Learning Discriminative Appearance Models for Online Multi-Object Tracking With Appearance Discriminability Measures

A data association, which links detections in consecutive frames, is a key issue in multi-object tracking (MOT). For robust data association in a complex scene, a common approach is to learn object appearance models for handling appearance variations of tracked objects and improving the discriminability between objects. However, learning appearances of multiple objects during tracking is still a challenging problem due to the frequent sample contamination by occlusions and low feature discriminability by similar appearances between objects. In this paper, in order to learn each object appearance, we propose a discriminative online appearance learning using a partial least square (PLS) method. In the proposed appearance learning, we first present online sampling mining to collect training samples from tracking results. Then, we consecutively learn PLS-based subspaces during tracking and discriminate object appearances by projecting object features onto the learned spaces. Since frequent appearance updates for all tracked objects increase the tracking complexity significantly, we propose measures to evaluate the discriminability of learned object appearances and update only the appearances with low discriminability. We apply the proposed appearance learning for online MOT and compare other appearance learning methods. In addition, we evaluate the performance of our MOT method on public MOT benchmark challenge datasets and show the competitive performance compared to other state-of-the-art batch and online tracking methods.