An Improved Framework for Pedestrian Tracking and Counting Based on DeepSORT

Due to the rapid growth in the demand for crowd analysis and monitoring methods, there is an urgent need for pedestrian tracking and counting methods that are more efficient, accurate, and reliable. However, the existing trackers and counters rarely have the three characteristics of efficiency, accuracy and reliability simultaneously. Therefore, we propose a framework named FR-DeepSORT for tracking and counting pedestrians based on DeepSORT. FR-DeepSORT first selects the YOLOv5 network as the object detector. Then the Re-ID information is combined with IoU to construct a cost matrix to improve the tracker by introducing FastReID. Finally, FR-DeepSORT introduces vector cross product and combines the tracking results to monitor the pedestrian crossing dynamics. The experimental results on the MOT tracking benchmark datasets show that the accuracy of our tracker is 89.38%, and the IDF1 value is 82.45%, which are competitive and more reliable to existing tracking methods. The counter error rate is also superior in the real-time dynamic statistical accuracy of the crowd.