Video Segmentation And Moving Object Extraction

Owing to the fast development of the computer techniques, it becomes not onlypossible but also necessary to access, manipulate, and retrieve the vast audio-visualinformation based on their contents. At the same time, new standards forcontent-based coding (MPEG-4) and description (MPEG-7) of audio-visualinformation are being developed or fulfilled. Thus, it becomes an urgent problem toextract and describe semantically meaningful moving objects in a video sequence. Asan important research area in video and image processing, the extraction of movingobjects in a video sequences has great value of research and application.The algorithms for extracting moving objects in video sequences are studied inthis paper. First, the moving object extraction system based on motion area detectionand spatio-temporal union are proposed. The main research work which have beendone in this thesis is described as the following:(1)In the aspect of temporal motion area detection, an improved high-orderstatistics algorithm is proposed to detect motion area. The experimental resultsindicate that this algorithm has the superiority of denoising compared with thetraditional high-order statistics algorithms. It can more accurately extract the motionarea from complex background.(2)Through applying the phase congruency edge detection algorithm based onLog-Gabor complex wavelet transform to the edge detection of frame image,one-pixel width edges of every frame are detected. For different kind of video images,this algorithm is compared with some classical algorithms in depth, such as Cannyedge detection and watershed segmentation etc. The experimental results show thatthe proposed algorithm is insensitive to tow contrast and illumination variation. Inaddition, the edge of objects detected is one-pixel width and is located accurately.(3)Combining improved high-order statistics algorithm with a phasecongruency edge detection algorithm based on Log-Gabor complex wavelet transform, a new algorithm for extracting moving object is proposed based on spatio-temporalunion information. The experimental results show that the proposed algorithm basedon spatio-temporal union information can detect moving object accurately from thevideo sequence with simple, complex background or with low contrast and seriousnoise.