Human Target Tracking Based On Scale Variation And Multi-Template Matching

The computing power of computer has improved greatly recent years,robot manufacturing technology and capabilities are mature.There are many new computer-based algorithms proposed,these algorithms and technology make the realization of robots with certain intelligence possible,especially they attract a lot of attention to improve the technology of service robot.For service robots,it is important to track the target effectively.A large number of video-based tracking methods have been proposed since 2012,and good results have been achieved.The tracking method is mainly applied to the security monitoring,however the tracking method in the actual scene is few,this paper uses the deep learning convolutional neural network method for image processing to achieve target tracking and make a solution for the application in the actual scenario.Aiming at the situation of target size scale change in target tracking,a special scale adaptation network is designed to extract the features of the corresponding scale,and effectively deal with the situation that the tracking object is overwhelmed by the background due to the change of distance.For the case where the target size is not considered in the video tracking,the target frame regression method is applied to scale the obtained target.In the case of the change of the target appearing in the actual tracking,this paper proposes an improved linear template updating method to enhance the robustness in long-term tracking.For the template that can be used in the template library,a comparison method is proposed,which combines the matching precision and the mean value of the previous three frames,so that the tracking history information is used in the template matching update,and the tracking algorithm is added to extract the target information.And the method in this paper is verified in the OTB data set and the actual environment.The performance of the algorithm is verified by the comparison test between the original algorithm and the algorithm of this paper.