Multi-target Detection And Tracking

Intelligent moving target detection and tracking system can significantly reduce theworkload to staff, improve work efficiency, and also can greatly improve the performance ofthe monitoring system. It is a hot research topic in the field of machine vision. Currently, ithas been used in intelligent transportation, health care, military and other fields, but theaccuracy and real-time problems are also existed in practical application.According to the requirement of the key project of national natural science foundation tothe multiple targets detection and tracking, a multiple robot detection and tracking system isset up based on network platform. Research about target detection and tracking algorithm isstudied on the experiment platform as well as testing. The main work is as follows:The overall hardware of the multiple target detection and tracking system has beendesigned and introduced in detail in this paper. After that, the hardware components in thesystem are introduced laying a foundation for software architecture. Software which controlthe hardware platform are introduced, such as VS2008, OpenCV image processing functionlibrary, etc. The overall chart designed in this paper is showed at the end.Research and testing of target detection algorithm is executed after designing thehardware and software architecture of system. Three target detection algorithms are studiedaiming at the static scene, including three main frame difference method, gaussian mixturemethod and codebook method. Filter processing, connected domain tagging and target fillingis used to dealing with binary image. The special threshold is applied to codebook method tospeed up the real-time performance of the algorithm. The three methods are successfullyapplied in this system. The improved codebook method has better real-time and accuracy todetect target in this system platform according to the experimental results.Target location must be known before target tracking, so the DLT of pinhole cameracalibration which is associated with target location is introduced at first. Then the method ofpositioning accuracy verification is introduced. Then the real-time low-pass filter isintroduced to fuse the target coordinate data of image overlap area, in order to improve thereal-time of image matching. Finally, the moving target tracking algorithm is studied ascamshift tracking, extended kalman filter. And camshift tracking algorithm is improved forgrayscale tracking focused on the condition of this experiment system. In addition, anon-linear discreted model of controlled object is established based on the classical model ofcarriage robot, which is a basic of the extended kalman filtering model.According to the experimental results and analysis about it, the improved codebook algorithm adding special threshold has good accuracy and real-time performance than othertwo kinds of algorithm in target detection. The extended kalman filtering algorithm has higheraccuracy than the improved camshift tracking in the target tracking. So the combination of theabove two methods is selected in target detection and tracking.