Moving Object Analysis Based On Computer Vision

Vision-based moving object analysis aims at detecting, identifying and tracking movingobject from video sequences, and furthermore understanding and analysis object activities.Related techniques have been broadly applied in video monitoring, robot technology, imageretrieve, and image compression and so on. The moving object detection and tracking is the basisof the object behavior understanding and analysis. Therefore it is of great significance to studymoving object detection and tracking based on vision. In this thesis, the moving object detectionand tracking are considered. The main research achievements of this thesis include:(1) According to the unsatisfied effect of target detection during the thresholdsegmentation process when background subtracted image takes on multi-peak characteristics,this paper presents a moving object detection algorithm combining background subtraction andedge detection. First, a new method is proposed based on advanced whitening response of greyprediction model. In the new algorithm, GM(1,1,x(1)) model, into which was introduced anadjustment parameter p, was adopted and only the pixels rounded the observation points wereselected as the model data. Then, by subtracting the obtained prediction image and originalimage, the edge detection was realized. Finally, the moving object detection is realized throughutilizing improved edge detection to detect the edge of object subtracted image.(2) This paper presents a two-stage object tracking method by combing region-basedmethod and contour-based method. In the first phase, Mean Shift combines with GM(1,1)prediction model to locate the original region of the object. This algorithm has optimized theinitial condition and background value of original GM(1,1) model whose prediction value isreplaced by that of improved GM(1,1) prediction model to reduce the iteration times of MeanShift algorithm and apply the method of multi-scale image information measurement to objecttracking in order to realize the adaptive change of tracking window; In the second phase, animproved level set method based on Mumford-Shah model is adopted to evolve the object ofoutline in order to obtain more precise object outline.(3) According to the disadvantages that the traditional Mean Shift algorithm has unsatisfied tracking effect in the conditions of complex environment, dynamic scenes, moving object orocclusion occurs, an tracking algorithm of moving object based imp roved Mean Shift andadaptive prediction is presented. The algorithm uses the features of both the color and gradient toimprove the Mean Shift algorithm, and then switches the filter by monitoring the Bhattacharyyacoefficient. When the object is in a normal movement, the initial iteration location of Mean Shiftalgorithm in next frame image is predicted by α-β-γ filter in order to reduce the iteration times;When the object is partially blocked or deviates in tracking, the particle filter is used to predict;When the object is completely occluded, the Mean Shift algorithm is discarded. Instead the α-β-γfilter is used to predict for maintaining the continuity of tracking. Experimental results show thatthe algorithm has high accuracy in tracking under static and dynamic scenes as well as the betterrobustness in the situations of irregular movement and severe blocks of the object.