Omnidirectional Vision Target Detection And Tracking Based On Embedded Platform

With the rapid development of monitoring technology,multi camera surveillance system with video analysis function has gradually become a new research focus.At present,most of the algorithms only stay at the theoretical research level,and the intelligent monitoring system for practical application is not satisfactory,which can not meet the needs of practical application.In view of this problem,this thesis studies the omni-directional vision target tracking system based on ARM embedded platform,and realizes the detection,recognition and tracking of multiple targets.After the deployment of the omni-directional vision system,the nodes will be static,and the background is relatively fixed,which is suitable for the target detection algorithm based on background modeling.Classical background modeling method is a hybrid Gauss model method,the detection effect is better but the operation speed is slow,it takes a lot of time and is easy to be affected by light.The model structure of VIBE algorithm is simple,small amount of calculation,fast detection speed,suitable for embedded platforms,but prone to ghosting and slow elimination.In this thesis,a local difference method is proposed to solve this problem.Experiments show that under the premise of ensuring the detection effect,VIBE formed in the first frame of the phantom of the opera can be quickly eliminated.Traditional tracking algorithms(such as particle filter algorithm)usually rely on color features,the detection effect is better,but the particles need to be updated,when the number of particles larger computing.CSK tracking algorithm based on correlation filter has good effect,but it can not realize the continuous tracking of the target is completely blocked.In this thesis,the Calman filter and CSK algorithm are combined to solve the problem effectively.When the object is completely occluded,the Kalman filter is used to track the target,and the CSK algorithm is used to keep track of the target when the target is out of the occluded area.The experimental results show that the target position can be predicted when the target is occluded completely,and the target can be tracked again.In this thesis,the camera is placed in a coaxial direction around the placement,the two cameras have a large overlap between the visual field,so it is related to the target at the time of tracking and the issue of the transfer of the same target matching problem.The target handoff problem,this thesis uses the method of vision line,the camera placed after reading the first frame image is corrected by SIFT algorithm,corner detection and adjacent camera,through the corner,generating vision line,and the resulting vision line back projection to the wide-angle camera.For the problem of target handover,this thesis uses a multi feature matching approach.Because the color feature is relatively stable,at the same time in pedestrian detection HOG features better,so firstly the color feature matching;when the matching target cannot be determined,using HOG feature matching,and achieve more stable target tracking.Finally,the algorithm is transplanted to the embedded platform,the processing ability of the embedded platform,and the memory has a large limitation,so it is necessary to optimize the algorithm effectively.With the help of multi thread,multi process,OPENGL and other means,the program architecture and implementation methods are optimized,and the algorithm efficiency is greatly improved.