Research On Multi-target Tracking Method In Video

In recent years,the theory of targets detection and tracking in video is maturing with further research. The main task is to deal with and analyze the targets in video image sequence,and to extract useful information from the targets,so that automatic recognition,prediction and ectraction of the targets location and tracking the movement of the targets. Targets detection and tracking in video is used in a very wide range of areas such as human-computer interaction, virtual realuty,medical diagnosis and intelligent traffic and so on. Therefore,the technology of targets detection and tracking in video has important research significance and application value.Nowdays, there are still many problems in the technology of targets detection and tracking in video, such as tracking on motion platform, change in the number of multi-target detection and tracking,target occlusion and so on.Due to those problems, targets detection and tracking in video is very diffucult and challenging.Multi-target detection and tracking under motion platform has been studied.Aim at the problem of tracking on motion platform,multi-target,long time detection and tracking drift and tracking failing caused by partial occlusion in complicated situations, a multi-target detection and tracking algorithm based on Haar-like detection and compressed feature fusion on motion platform is proposed.The algorithm is focused on Haar-like feature extraction of the moving targets, target location extraction, compressed features extraction and compressed feature fusion.The paper dicides into two parts:target detection and target tracking.In the target detection,some classical methods,such as frame difference method, background subtraction method,motion estimation method under static background and HOG feature detection,LBP feature detection,Haar-like feature detection under dynamic background,are introduced.And the advantages and disadvantages of several methods are compared.In this paper,face image database and vehicles image database for training target detection classifier are built.Haar-like feature detection is used. The experimental results show that this method can quickly and accurately detect multi-target candidate regions in video images sequence,and the detection results laid a good foundation for subsequent target tracking.In the target tracking, some mainstream methods,such as LK optical flow method, Meanshift, TLD and CT, are introduced.A algorithm is proposed based on compressive tracking.According to the compressive sensing theory, future dimension of target is reduced in the part of feature extraction process.A tracking method based on compressed feature fusion is proposed.When partial occlusion occurs, this method can significantly reduce the loss of image feature information,and reduce tracking drift even tracking failing..Experiment results show that the proposed method ensures tracking stability and robustness.