Distractor-aware discrimination learning for online multiple object tracking

Zongwei Zhou, Wenhan Luo, Qiang Wang, Junliang Xing, Weiming Hu

Abstract: Highlights • A distractor-aware discrimination learning model is proposed to facilitate online multi-object tracking to better differentiate one target from other targets and semantic backgrounds in the scenes. • A relational attention learning mechanism is introduced to handle appearance variations of targets caused by large pose variations, object occlusions, and target interactions. • A multi-stage tracking strategy is established within a temporal sliding window which leverages the object detection responses and tracker predictions to deal with trajectory drifting. • Extensive experimental analyses and evaluations on the widely used challenging MOT16 and MOT17 benchmarks demonstrate the effectiveness of the proposed approach. Abstract Online multi-object tracking needs to overcome the intrinsic detector deficiencies, e.g., missing detections, false alarms, and inaccurate detection responses, to grow multiple object trajectories without using future information. Various distractions exist during this growing process like background clutters, similar targets, and occlusions, which present a great challenge. We in this work propose a method for learning a distractor-aware discriminative model that can handle continuous missed and inaccurate detection problems due to the occlusion or the motion blur. To deal with target appearance variations, a relational attention learning mechanism is proposed to capture the distinctive target appearances by selectively aggregating features from history states with weights extracted from their appearance topological relationship. Based on the discrimination model, a multi-stage tracking pipeline is designed for automatic trajectory initialization,propagation, and termination. Extensive experimental analyses and comparisons demonstrate its state-of-the-art performance on widely used challenging MOT16 and MOT17 benchmarks. The source code of this work is released to facilitate further studies on the multi-object tracking problem.1