Research on pedestrian tracking algorithm based on deep learning framework

Abstract

Pedestrian tracking is a hot topic in the field of computer vision. Current pedestrian tracking methods may treat the reappear target as a new target to tracking, which can lead to tracking failure. For this purpose, this paper uses the deep learning framework to complete the whole algorithm of pedestrian tracking to solve this problem. The pedestrian tracking algorithm is divided into pedestrian detection and tracking. In the detection part, the Faster-RCNN framework is used to detect pedestrians; and in the tracking part, we use the Person-ReID method to track pedestrians, converted the pedestrians tracking problem to feature extraction and matching problem between different frames, and improved the tracking results effectively according to the deep learning framework. According to the experiment results on the simple standard dataset and RGB-D People dataset, the tracking mAP of our algorithm get 92.51% and 76.9%.