Research And Improvement Of Kernel Correlation Filter Tracking Algorithm On DSP Platform

With the development of machine learning and deep learning,high-performance intelligent target tracking technology has gained the attention of researchers and is widely used in various fields.In recent years,the KCF tracking algorithm,which have high tracking accuracy and can also achieve real-time tracking,have occupied a place in practical engineering applications.With low computational complexity,the algorithm is also suitable for deployment on embedded platforms with limited hardware resources.However,the KCF tracking algorithm performs poorly in scenarios with target deformation,scale change,and target occlusion,and tracking failure can easily occur.Therefore,it is highly valuable to investigate tracking algorithms with higher robustness in embedded platforms for engineering applications.In this thesis,the KCF algorithm has been studied and improved,the main works are as follows.(1)To address the problem that KCF algorithm cannot adapt to target deformation or target scale changes,this thesis proposes a scale adaptive algorithm based on feature fusion,which fuses the FHOG + ULBP-ROT + CN feature,and the fused features can describe the target in an all-round way so as to better overcome the impact caused by target deformation.In order to solve the problem that the KCF algorithm cannot adapt to the scale change of the target,after the tracking algorithm finds the target,a multi-scale sampling of the target is performed and a scale filter is trained to determine the size of the tracking frame,thus reduce the tracking failure rate caused by scale change of the target in the tracking process.(2)In order to improve the anti-occlusion performance of the KCF algorithm,a dynamic search algorithm based on model self-adaptation is proposed in this thesis.The scheme adopts an occlusion discrimination index based on confidence to judge the extent of the occlusion,and adopts a corresponding model updating strategy through the tracking state after discrimination to achieve the purpose of model self-adaptation,the model updating strategy effectively prevent the degradation of the model.Besides,in the case of severe occlusion,dynamic search will be performed to raise the success rate of target relocation.(3)In order to improve the limitation of the search range of the dynamic search algorithm based on model self-adaptation,this thesis proposes a relocation algorithm based on rough and fine match,which first locates the approximate position of the target in the global range by rough match,and then uses fine match to correct the rough match results,so that the target can be accurately located.The algorithm is used in conjunction with the dynamic search algorithm based on model self-adaptation,which has certain advantages over the traditional occlusion processing algorithm in terms of searching and relocating the target after its reappearance,and effectively improves the anti-occlusion performance of the algorithm.(4)All the improved algorithms based on KCF algorithm are fused into HRKCF algorithm,the HRKCF algorithm is transplanted to the DSP side,and the algorithm is optimized by the algorithm library which is provided by the DSP platform,not only reduces the amount of code,but also the speed of operation has been improved.In addition,the software framework of the embedded target tracking system based on FPGA + DSP is designed and implemented.The system mainly includes SRIO-based image transmission,upper computer interaction based on RS422,image display,target tracking and other functions.Besides,in order to reduce the coupling degrees of all modules,the software framework is improved by using SYS/BIOS framework in a multi-threaded way.