JDE multi-object tracking algorithm integrating multi-level semantic enhancement

In order to solve the problem of target ID switching caused by occlusion and insufficient information location extraction in JDE(joint detection embedding) algorithm, an improved multi-target tracking algorithm based on JDE is proposed this paper. Firstly, SPA feature space pyramid attention module used expand receptive field obtain more abundant semantic improve accuracy model for different scale targets. Secondly, FCN network makes header Embedding task collaborative learning alleviate excessive competition enhance original information, effectively reducing number switching. Finally, PCCs-Ma motion measurement can strengthen connection between Kalman filtering prediction observation, reliability similarity discrimination characteristics. verify effectiveness are compared same experimental environment. The results show that average 3.94 %. On MOT16 dataset, MOTA IDF1 indexes increased 6.9 %, significantly reduced, good achieved.