Research On Real-time Target Tracking Algorithm And Optimization Method Based On Embedded Platform

As one of the basic technologies in the field of computer vision,target tracking technology has a wide range of applications in automatic driving,video surveillance,weapon guidance and so on.Among the embedded platforms with various architectures,the embedded platform with the architecture of FPGA and DSP has advantages in image processing operation and has high cost performance.Therefore,the embedded platform based on the architecture of FPGA and DSP is chosen to complete the task of target tracking.According to the weak computing power of the platform and the requirements of different application scenarios,the tracking algorithm based on auxiliary positioning in local area and the target tracking algorithm based on re-detection in the whole image were proposed and transplanted to the embedded platform.The specific work carried out in this paper is as follows.1.The tracking algorithm based on auxiliary positioning in local area was proposed.The algorithm consists of two parts: the self-adaptively updated spatiotemporal context continuous tracking part and the self-adaptively updated compressed sensing assistant localization part.The OTB2013 standard test framework is an international general target tracking algorithm evaluation framework,which can evaluate the comprehensive performance of the algorithm in complex tracking scenarios.This framework was used to test the tracking performance of the algorithm.The tracking accuracy of the algorithm was 13.35% higher than that of the benchmark algorithm,and the success rate was 27.46% higher.The algorithm was transplanted to the embedded system platform,and real-time stable tracking was realized.2.The tracking algorithm based on re-detection in the whole image was proposed,which can re-lock the target after it went out of the field of view and returned from any path.The algorithm includes two parts: self-adaptively updated spatio-temporal context main part and the normalized cross-correlation target detection part.The performance of the algorithm was tested using OTB2013 standard test framework.The accuracy of the algorithm was 5.19% higher than that of the benchmark algorithm,and the success rate was 15.90% higher.After algorithm transplantation,the target-out-of-view scene was created manually.When the target re-enters the field of view,the tracking algorithm based on re-detection in the whole image can accurately lock the target and continue to track stably.3.In the process of algorithm implementation in the embedded platform,in order to improve the efficiency of algorithm operation,this paper optimized the program from the algorithm and the characteristics of the embedded platform.The program was optimized in program structure,compiling environment,software pipelining,memory space,support library.After optimization,the algorithm can realize real-time tracking in the embedded platform used in this paper.The frame rate of the tracking algorithm based on auxiliary positioning in local area and the frame rate of the tracking algorithm based on re-detection in the whole image were 53 frames per second.The optimization effect is significant.