Stable and real-time pedestrian tracking method based on improved DeepSORT under complex background

The real-time multiple object tracking algorithms have achieved ideal tracking performance, but the tracking speed is slow, and tracking accuracy is also decreased with the increase of background complexity in most of the recent algorithms. In terms of these issues, a real-time multiple pedestrian tracking algorithm is proposed based on online data association. First of all, the dual prediction mechanism of kernelized correlation filter and Kalman filter is designed. This mechanism forms a prediction tracking calibration system with the cascade matching in DeepSORT, which makes the data correlation more reliable. In addition, the attention mechanism is introduced in the object detection part of tracking to enhance feature representation ability by strengthening position information of the object, so as to improve racking accuracy. The experiment is carried out on MOT16 dataset, the MOTA is up to 66.5%, IDF1 is up to 64.2, IDSW is 641. Compared with DeepSORT algorithm, MOTA and IDF1 increase 13% and 13.2% respectively, and IDSW decreases 410. Experimental results show that the proposed algorithm is helpful to solve the problem of object false detection, missing detection and other problems in multiple pedestrian real-time tracking. It still maintains high tracking accuracy for severe occlusion in tracking, which can achieve real-time and stable multiple pedestrian tracking in complex background.