Research On The Algorithm For Moving Target Detection And Track In Extremely Wide View Through A Fisheye Lens

Because of the increasing demands of computer vision application, Omni-directionalVision has been widely concerned and researched in recent years. The traditionalOmni-directional vision system is based on Omni-directional Vision Sensor (ODVS). ODVScan capture3600Field of View image in horizontal direction with one time shooting, whichis based on some kinds of reflection mirror. We can easily find that there exists a central blindarea or some useless information such as camera or camera stand in the acquired image.Fisheye lens has been used to build the Omni-directional Vision system in this paper, whichhas the advantages such as no blind area, without images joining, with abundance informationand so on. The widely predictable applications of fisheye lens are including: SecuritySurveillance, Pipeline Inspection, Assistant Driving, Field Monitoring, Aerocraft Guidance,Space Robot etc.In the thesis, the research mainly concerns about moving objects detecting and trackingalgorithms and methods under the fisheye lens with extremely field of view.Some of classical objects detecting algorithms have been reviewed firstly. There arebackground subtraction, inter-frame differential method, and Optical flow. After that, wepropose an objects detecting algorithm which integrate Gaussian background modeling,inter-frame differential method, and Optical flow into one. Firstly, we use Gaussianbackground modeling to update the background in real time. Then background subtraction hasbeen applied to get a moving target with the current frame and the background. Next, we canget another target by three frame difference method. Finally two moving targets can be gotwith the OR operation. Experiments show that the improved algorithm has good robustnessand real-time performance, which can be used for detecting the moving objects under thefisheye lens images.The other part of this paper is about moving tracking research. Mean-shift algorithm andCamshift algorithm has been reviewed firstly. The relationship between Mean-shift andCamshift has been discussed. Camshift algorithm can adapt to dynamic probabilisticdistribution and have the capability to deal with the continuous image, so it is used to makethe basic framework of the new tracking algorithm. In order to improve the trackingperformance for fisheye images, we integrate the Kalman Filter into Camshift algorithm toforecast the next position of the moving targets. Experiments prove that the improvedCamshift algorithm combined with Kalman Filter has good robustness and real-timeperformance, which can satisfy the demands of moving objects tracking under fisheye lensimage.And the improved algorithms are applied to the field of Omni-Vision Video Surveillanceand the navigation of AGV. The improved algorithm of moving target tracking has a good performance in AGV navigation with good robustness.