MPTS-AFBP: Multi-pedestrian Tracking and Segmentation Based on Anchor-Free Detector

In the last two years, a small amount of work has explored a new task called Multi-Object Tracking and Segmentation (MOTS), which aims to classify, locate, segment, and track all instances of a particular category in the whole video sequence. However, the existing MOTS algorithms have some problems, such as low segmentation accuracy and frequent object identity switching. In order to solve these problems, we propose a multi-pedestrian tracking and segmentation method based on anchor-free detector. Subsequently, in the segmentation branch, we fuse multiple information to refine the mask through blender module. In addition, in the tracking branch, we encode the position information of the object and combine it with appearance feature to generate association vector. In order to test the effectiveness of our method, we have carried out verification on the KITTI MOTS and MOTS Challenge dataset, and the experimental results prove the effectiveness of the method.