| 英文摘要 | In recent years video surveillance technique is popularly used, and based on it intellegent video analysis algorithms attract many researchers and many useful algorithms have been proposed in the literature, among which the most basic algorithms are that of moving object detection and tracking. However, many object tracking researches focus on single camera, which has limited field of view.Though multicamera system is popular used nowadays for obtaining large field informations, few up-to-date researches have been proposed to address object correlations between multi cameras. Multi videos generated by multicameras contain indevidual visual information of the corresponding surveillanced field of view, thus by smartly linking moving objects in different time and locations, plenty information can be obtained for security surveillance, which is more powerful than just viewing two cameras seperatly. This is very important for future applications of the Internet Of Things. Therefore, researchers pay more and more attentions on hand-off object tracking between multi cameras. However, it is just at the begining stage, and there exist many problems to be solved, such as problems caused by complex illumination, viewpoint, pose, and distance variations between multi cameras, the diffculty of having many video data to be labeled, the problem of model adaptation in new camera environment, and so on. This thesis focuses on continiously tracking multi objects between multi cameras, addressing many basic problems in image video analysis and pattern recognition, including illumination invariant local feature description and selection, object detection and segmation in complex scenes, online learning of illumination, view point, pose, and distance independent features for multicamera tracking, online learning of space and time constraint for multicamera learning, and so on. The main works and contributions of this thesis are as follows. The thesis proposes three effective local texture descriptors: Structured Ordinal Feature, Block-based Local Binary Pattern, and Scale Invariant Local Ternery Pattern, achieving robust performances against illumination variations, and they are evaluated with the effectiveness for object detection, recognition and tracking. The thesis proposes a moving object detection approach for complex scenes, which uses the proposed scale invariant local ternery pattern for the description of complex background under illuminati... |
修改评论