Research And Implementation Of Intelligent Object Tracking Algorithm Based On Embedded Platform

Visual object tracking technology is one of the research hotspots in the field of computer vision,Object tracking technology is widely used in military and defense security fields such as navigation,guidance,drones and surveillance.In addition,it is also widely used in civil fields such as intelligent traffic monitoring and security.With the vigorous development of artificial intelligence and computational vision technology,the object tracking field has also made great achievements,and many high-precision intelligent object tracking algorithms based on machine learning methods have been proposed one after another.In engineering applications,the tracking algorithm of template matching is still widely used.Such algorithms can not handle occlusion,illumination and other changes well.Therefore,the intelligent object tracking algorithm is studied and transplanted and applied to the actual embedded.That is important theoretical research value and engineering application value in object tracking system.In recent years,object tracking algorithms based on correlation filtering have developed rapidly in the field of visual tracking,and have achieved great success in tracking accuracy and real-time performance.The kernelized correlation filter tracking algorithm skillfully utilizes the use of cyclic shift to generate a large number of training samples,and utilizes the nature of the diagonalization of the circulant matrix,so that the algorithm is mainly calculated as FFT fast Fourier transform.However,the original kernelized correlation filter tracking does not judge the confidence of the tracking result,which makes the target easy to track down in the occlusion scenario.The main research contents of this paper are as follows:(1)Firstly,the principle of the kernelized correaltion filter tracking algorithm is studied,and analysis the advantages and disadvantages of the algorithm.To solve the model update problem in the case of correlation filter occlusion,this paper introduces the SVM appearance model based on the APEC high confidence,combine the two to make a confidence judgment on the target tracking result.After that,the improved algorithm performs a large number of tests on the actual scene video sequence.Experiments show that the improved kernelizes correaltion filter tracking algorithm improves the occlusion performance on the basis of realtime performance.(2)This paper implements an improved kernelized correlation filter tracking algorithm on the TMS320C6657 DSP platform.Among them,the image two-dimensional FFT calculation,image HOG feature extraction,SVM transplantation and other functions are realized on the DSP.On the basis of the transplantation implementation,a large amount of simulation and debugging is performed on the DSP platform using the actual scene video sequence.The simulation results show that the tracking algorithm transplanted to the DSP side is basically the same as that on the PC,and the tracking algorithm after transplantation is only 10 ms in the DSP platform,and the tracking frame rate is up to 100 fps.(3)Based on the algorithm migration of DSP platform,this paper implements the software framework design and development of real-time target tracking system based on FPGA+DSP architecture,including video transmission,PC host computer interaction,tracking result display.The video transmission process between FPGA and DSP based on SRIO interface and the analysis and execution process of DSP interactive instruction based on RS422 serial port are introduced in detail.Finally,based on the TI SYS/BIOS framework,the modularization and multi-threading design and implementation tasks of the software system are completed.