Online Multi-target Tracking Algorithm In Crowd

The task of video target tracking is to process and analyze the target appearing in the video to find the location of the target and the motion track.It is a key technology in the field of computer vision research and is widely used in intelligent transportation,intelligent security monitoring,and fields such as human-computer interaction.Therefore,the research on video target tracking algorithm has important application value and broad development prospects.In the target tracking research,online real-time multi-target tracking is the most widely used and most difficult one.With the increase of outdoor peoples number,the demand for tracking algorithms for crowded scenarios in growing too.In this thesis,we chose the competitive algorithm deep SORT to study the ID switch and the running speed of the crowded scene,and made the following improvements:1.When the current detection bounding-box is matched with the Kalman predicted trajectory position,for the lower confidence matches,use the extra distance metric to match again,so that the ID switch is reduced with the least time cost to increase the tracking in the crowded scene.2.In the Kalman filter application of the multi object tracking algorithm,the Kalman gain is automatically adjusted when the targets occlude higher than a certain threshold,and the number of ID switch occurring during pedestrian interaction is reduced with the least time cost.3.In the process of using the Hungarian algorithm,the Hungarian algorithm is improved for the cost matrix feature in the crowded pedestrian scenario,so that the step of finding the optimal matching solution is reduced.We implemented the above improved algorithm and tested it in the crowded scene data set MOT16 of MOT Challange.The test results are as follows: other performance indicators are basically unchanged,and the indicator IDs representing the robustness of the track ID are reduced by 8%.It can also reduce the number of matching optimization iterations.