关键词: 多目标跟踪, 动态模型, 特征匹配, 人头跟踪, 行人重识别

Abstract: Substantial improvement has been achieved in the field of Multi-Object-Tracking due to the successful application of deep neural networks. However, tracking pedestrians in crowd scene remains a challenging problem in computer vision. Previous methods work well in common scenes but failed in the crowded situation since severe occlusion between pedestrians that makes the motion model and appearance feature extraction failed. Inspired by the less occluded head part, we propose a part-based motion and Re-ID model which makes full use of both the head part and full body for tracking. There are mainly two modules in the proposed method. Firstly, instead of using full body bounding box for motion modeling, we use the head part to model the pedestrian position since the head part has few occlusions. Secondly, in order to get denoised full-body appearance feature in crowded scene, we propose an adaptive full-body bounding box generator to obtain full-body bounding box from head bounding box and use the human pose to guide the Re-ID model to extract appearance features. Our model achieves the state-of-the-art result on Head Tracking 21 dataset. We show the necessity of each module of our approach with extensive ablative study. Moreover, by adaptively generating full-body bounding box, our approach could also be used for full-body tracking.

Key words: Multi-object-tracking, Motion Model, Feature Matching, Head Tracking, Occluded Person Re-ID