Design And Implementation Of Multi-target Tracking System Based On Embedded Platform

With the rapid development of science and technology and the improvement of social life,more and more high-tech products are used in life and become an indispensable part.In the field of security protection,video surveil ance plays a decisive role.In intel igent video surveil ance,multi-target tracking technology is a research hotspot.This technology provides important information such as target position and trajectory of interest in video,and is a prerequisite for subsequent target recognition and behavior analysis.However,in the actual scenario,we need to control factors such as power consumption,manpower and material costs.This article is devoted to designing a set of unmanned,real-time and efficient multi-target tracking system.Firstly,in this paper,for the application of intelligent video surveillance and the needs of complex scenarios,it overcomes the long-term problem of multi-target tracking algorithms relying on offline detection,and designs and develops a multitarget tracking system based on online detection.The embedded TMS320C6678 DSP core board,Kintex-7 FPGA backplane,and Jetson TX2 development board are used to collect images and run algorithms in real time.Combined with the UI interface of the upper computer,it forms a complex real-time multi-objective system,which has the advantages of automatic and flexible,real-time online tracking and so on.Secondly,In the multi-target tracking algorithm,based on the requirements of the application scenarios and the defects of the traditional algorithms in occlusion,low real-time,etc.,based on the SORT multi-target tracking algorithm,four-point improvement strategies are proposed to achieve stable tracking of the target trajectory.First,the target is characterized by the combination of apparent information and Kalmanian motion.Second,the cascade matching strategy is introduced,and the nearest first match principle is adopted to solve the problem of trajectory al ocation errors due to camera movement and long time occlusion.The third problem is that the clustering algorithm is used to adapt to the crowded scene and solve the problem that the trajectory processing speed is reduced due to the excessive number of targets.Fourth,because the multi-target tracking accuracy depends on the performance of the detector,in order to improve the algorithm in complex scenes.In the tracking accuracy rate,this article uses the YoloV2 detection algorithm as the input of the multi-target tracking algorithm,and adds a suppression strategy for false alarm detection information.Finally,in the system software construction,is divided into DSP software,FPGA software and TX2 software in three parts.The DSP software runs a multi-target tracking algorithm with the SYS/BIOS real-time operating system as its core.The software uses multi-threaded design ideas and C/C++ mixed compilation to make the code more concise and efficient.At the same time,OpenMP parallel methods are introduced to increase development and parallel efficiency.The interface system establishes the TCP/IP protocol stack and uses TCP communication to interconnect devices,providing stable data links.In the FPGA software,it is responsible for image acquisition and communication with the DSP through SRIO to ensure the system's effective image input reliability.In the TX2 software,the YoloV2 detection algorithm is run to provide input to the tracking system.In addition,the upper layer software developed with the MFC framework has been added to provide an interface display of the tracking system.The multi-target tracking system designed by this article is composed of a standard-definition camera,K7 FPGA front-end acquisition,dual-chip DSPC6678 realtime tracking operation,TX2 detection platform and host computer interface.At the end of this paper,the system and algorithm will be tested and analyzed to verify the integrity and real-time performance of the system.