Research On Human Body Multi-Target Tracking

With the development of science and technology,multi-target tracking technology in video has always been an important research content of many artificial intelligence research scholars.Human body multi-target tracking is widely used in many fields,such as video surveillance,driverless and human motion analysis and so on.Although many researchers have made great achievements in multi-target tracking,there are still some problems that need to be solved due to the influence of many factors such as illumination,human body staggering motion,occlusion,and deformation,etc.The multi-target tracking effect is also closely related to the target detection.For the influence of different factors in the tracking process,a more optimized method should be adopted to avoid or solve.For different influence factors,the main research contents and innovations of this paper are as follows:(1)Aiming at the complex problem that the SSD target detection algorithm needs to implement the multi-box prediction for each grid in the feature map during the detection process,an improved method to avoid redundant calculation is proposed.Firstly,the background in the field of view is used for background modeling,and regions where targets appear in the current video frame are obtained by using the self-organizing background subtraction method,and then the SSD is used to classify target region boxes to obtain the detected human targets,thereby avoiding multi-box detection on these grid cells that only contain backgrounds.For the problem that SSD detection algorithm has poor effect on small target detection,a method for implementing feature fusion using context information is proposed.(2)Aiming at the regional density of targets and the real-time tracking problem in multi-target tracking process,a multi-target tracking algorithm based on region partitioning is proposed.Firstly,the target obtained by the detection is used as the node of the graph,and the correlation clustering algorithm is used to divide the region.Then,the correlation between the detection target and the tracking target is used to classify the tracking target into the divided region.Finally the tracking scene isdivided into simple tracking regions and complex tracking regions.In the simple tracking region,simple color feature is used as the tracking basis;In the complex tracking region,the tracking model with complex features is used to achieve the target tracking.(3)Aiming at the problem of occlusion in complex tracking regions and the effect of fast motion on target tracking,a tracking method based on local feature is proposed.Firstly,the Kalman filter is used to predict the motion range of the occlusion target.Secondly,tracking targets are meshed,and a tracking model based on local feature adaptive weighting is trained by the improved convolutional neural network(CNN),the offline kernel correlation filter tracking algorithm is combined to implement target tracking under local occlusion and full occlusion.Finally,the experimental results on the MOT-2016 dataset show that the proposed human body multi-target tracking algorithm can better deal with the tracking problems existing in complex scenes and has better real-time performance.