Multi-Object Tracking with Multiple Cues and Switcher-Aware Classification
Weitao Feng, Baopu Li, Wanli Ouyang
Abstract
Multi-Object Tracking (MOT) has been a popular and challenging topic in computer vision. However, identity issue, i.e., an object is wrongly associated with another object of a different identity, still remains to be a difficult problem. To address it, two factors are of great importance. First, multiple cues of different sources are needed for robust tracking to handle complicated situations where single source cue may not be reliable. Second, switchers that confuse targets and cause identity issues should be paid more attention to provide more information and avoid such issues. Based on these motivations, we propose a method for MOT, which mainly aims to take more cues and information of potential switchers into consideration. Other than the frequent usage of single appearance cue, we exploit cues from tracklet surroundings and historical appearance features and combine all cues in a unified manner. Unlike usual tracking methods, the proposed tracking classifier learns to deal with different strategies in varied situations w.r.t. a switcher. Extensive experiments show that our proposed method achieves competitive results in the challenging MOT benchmarks.