Research On Visual Target Tracking Algorithm Based On Verification And Error Correction Strategy

With the arrival of Safety City,the video surveillance field has developed rapidly,and various intelligent terminals have emerged.The visual target tracking as a key technology of video surveillance has played an important role.As the visual target tracking technology can be applied to the automatic driving system to track the targets,which has drawn much attention from researchers.Although there are many excellent visual tracking algorithms,these algorithms can not meet the requirements of speed and accuracy in real-world scenarios.Aiming at the real-time and accurate requirements of the visual target tracking algorithm in real-life scenarios,this study will deeply analyze the existing domestic and foreign classical visual target tracking algorithms,and put forward some efficient solutions to address some key problems about algorithms.Finally,an flexible,efficient and highly accurate visual target tracking algorithm,named ATVC is proposed.At first,this paper analyzes the principles of ECO and SINT algorithms.;Second,we find some key questions in these algorithms;Third,we put forward some corresponding solutions to these specific problems.For the target to be tracked in the video,some targets are easy to track,while others are difficult.We can adopt different solutions to deal with according to its difficulty degree of the video frames.In order to distinguish the difficulty degree related to the video frame,this paper designs a new verification module to solve this problem;For the problems of target re-tracking and re-emergence in the visual tracking algorithm,this paper proposes a flexible and efficient error correction module,which can be embedded in any tracker.In term of the SINT algorithm,this paper proposes several new improvements to improve the tracking performance of the algorithm.We introduce the ROIAlign layer to obtain a more accuracy bounding box;For the purpose of gaining a more robust feature representation,the residual blocks is added to.At the same time,the speed of inference at improved algorithm is shortened;So in order to integrate the problem of scale and appearance change in the visual target tracking algorithm,we try to incorporate the idea of FPN,and it enhances the robustness of the feature dramatically.For the sake of verifying the effectiveness of the improved visual tracking algorithm,this paper tests the performance of several classic tracking data sets such as OTB50,OTB100,TC128 and UAV123.In order to evaluate the performance of the proposed algorithm,this paper conducts an objective index evaluation and subjective index evaluation in different data sets.In addition,in order to prove the effectiveness of the improved scheme proposed in this paper,we test the performance improvements from different modules respectively.The results of a large number of tests show that the proposed visual target tracking algorithm not only achieves higher precision,but it is also more flexible,and can meet the needs of real-world scenarios.Furthermore,it can tailor the module for impossible tasks.The visual tracking algorithm proposed in this paper can obtain a frame rate of 35 fps on the GTX 1080 GPU,which basically meets the needs of real-time scenes.In summary,the visual target tracking algorithm ATVC,which proposed in this paper has many advantages such as high precision,ease of use,high flexibility,and the meeting the needs of real-world scenes.In addition,it can be tailored to specific tasks in real-life scenarios,and can meet the demand of the realistic scenarios.