Multi-target Tracking Method Based On Deep Learning In Complex Scenes

In the field of computer vision,multi-target tracking is a hot research direction,which has great academic value.For example,the intelligent video system of full-automatic,all-weather and real-time monitoring in smart city.In the traditional video surveillance system,for multi-target tracking,there are many problems,such as consuming a lot of human resources and can not be processed in real time.The multi-target tracking technology based on deep learning is an efficient end-to-end learning framework,and the deep neural network model is an effective method to capture the information strongly related to the target tracking position.In the actual complex scene,such as high crowd density,serious occlusion area,and facing the background interference problem,the effect of multi-target tracking trajectory is still unsatisfactory.In order to improve the effect of trajectory prediction,this thesis proposes a multi-target candidate region feature extraction method based on deep learning and a hierarchical Association strategy combined with trajectory confidence,and establishes a hierarchical association model of multi-target prediction trajectory based on residual network.The specific methods are as follows: firstly,the tracking target motion is predicted by Kalman filter,and then the candidate box is corrected by the target detection method based on deep learning,so as to improve the tracking accuracy of the target location.Then the tracking targets are classified accurately by classification network.Among them,when dealing with the targets with similar features,the deep hidden features of each target are generated based on the neural network,and then the most similar trajectories are linked.By distinguishing two overlapping and staggered candidate trajectories,the purpose of reducing the loss of trajectories and maintaining the stability of tracking is achieved.Finally,the track generated by the tracking target in different frames is accurately matched by hierarchical data association.The experimental results show that the proposed model is better than the existing methods in dealing with the problem of false detection and improving the recognition accuracy.