Research On The Key Technology Of Video Multi-target Tracking Based On Tracklets

The age of artificial intelligence has come to us.The intelligent detection and analysis of video have become an important area in computer vision.As an important research topic in video analysis,target tracking plays a very important role in traffic monitoring,traffic flow statistics and so on.The main task of multi-target tracking is to determine the position of each observation target in the video sequence,and to mark the same target number in the same target area to form the trajectory.However,there are many factors that affect the tracking effect in the complex motion scene and background,including the occlusion,the deformation produced by the motion,the similarity of the adjacent targets and so on.To solve the problems,we propose a new multi-target tracking method based on tracklets association.The main research contents are as follows:1.Research on detection method and data preprocessing.By learning the mainstream target detection algorithms,the Deformable Part Model(DPM)algorithm is used to detect the multiple targets.In the process of preprocessing the detection result,the non maximum suppression method is used to remove the repeated detection blocks.Meanwhile,the adjacent target frames are searched to find the neighboring target blocks.And size comparison method is used to judge whether the target block is wrongly detected,thus eliminating the error detection targets and providing more accurate inputs for the tracking task.2.A high confident tracklet generation method.We use SIFT to extract feature points in the target blocks.The reverse optical flow tracking method is adopted to track the feature points in the target blocks,and the inaccurate target blocks are corrected by the matching results of the feature points and comparing the size of the target blocks.Finally,the target association is realized through the effective combination of the target's light flow,location,color and size,thus generating high confidence tracklets.3.An improved network flow model with multi-similarity fusion method.In this paper,the network flow graph of the tracklet is constructed in the way which every tracklet is recognized as the node in the network flow graph.And the edges of the nodes are established in the order of time.A new color feature based on super-pixels and DPM is proposed.Motion estimation is used to predict the target position.All of the measure indexes including the color similarity,position correlation and trajectory smoothness are used to construct the cost of the edges in the network flow graph.Finally,shortest path algorithm is adopted to find the all minimum cost flows to get the multi-target trajectories.The final experimental results show that,this method has obvious performance improvement compared with some multi target tracking methods proposed earlier,which highlights the value and significance of this study.