Moving Object Detection And Recognition In Video Sequence

The thesis is focused on moving target detection, segmentation and recognition in video sequence. As a widely used technology, video analysis is task-dependent. The thesis tries to find the intrinsic connections among different algorithms, which are employed in different applications. Then, with deeper understanding of them, we hope to make some improvements.In the aspect of object detection and segmentation, we begin with the static background case. Two modified algorithms, adaptive time-delayed difference model and mixture background estimation model based on MRA, are proposed. Considering the weakness of the Optic Flow approach, we give an improved object detection and segmentation method in the moving background case. Afterwards, two practical factors, which bring negative effects, are discussed in detail: lighting variation and camera dithering. Since the automatic algorithm can handle very limited real situations, the interactive (or semi-automatic) approach seems to be an effective alternative. Based on 3D MeanShift segmentation, we propose an interactive video retrieval method.Next, we explore similarities and differences of target recognition between moving target and target in the static image, and find the appearance of moving target is time varied. Therefore, a multi-view sampling feature extraction approach is given. Experimental results show it works well for complicated motion target recognition.In the end, we discuss the design of a prototype system facing to video analysis. And, a structural representation in the framework of XML is produced for the result of video object detection, segmentation and recognition.