Video-Based Detection Of Abnormal Behavior In The Examination Room

With the rapid development of computer and network technology, as well as the promotion of multi-media applications, the amount of digital video has grown exponentially in recent years. Ensuing question is how to more effectively search the information of interest in the massive video database. A key part of this system is the ability to search over vast amounts of video data. While traditional methods have relied on text, such as those extracted from closed captioning, speech analysis, or manual annotation, we would like to search based on the automated recognition of the visual events in the video. The current video monitoring technology in examination room is still used manual observation and recording methods, not only time-consuming, prone to missed, but also judge the outcome of existing defects such as subjectivity and inaccuracy. We propose a method for visual behavior detection of human actions that occur in crowded, dynamic environments. In this paper, main research work includes:(1) Aiming at the requirement of video analysis, this paper presents the abnormal behavior detection method. Design of the software and hardware system and related settings, then we propose an effective and reasonable scheme for examination room video capture. Aiming at the partial occlusion and background clutter in the examination room, we compare several video segmentation methods. And then we propose a spatial-temporal video segment technology, which has achieved good results.(2) Studying the technology of Interactive Video Segmentation, and extract he 3D spatial-temporal template. Interactive segmentation method can make accurate segmentation boundary with manual intervention. Experiments show that this method performs well in our datasets. Study the unsupervised clustering method based on the test video over-segmentation. Aim at the slow implementation of the Mean Shift algorithm, we use octrees to perform an initial coarse clustering of voxels that are similar in color. And then we perform hierarchical clustering using Mean Shift that segments the video. The experimental results show that this method enables a factor of 100-1000 speedup in computation time without significant loss of quality. (3) Study the volumetric shape matching. Aiming at shape-based correlation algorithm can sometimes generate false positives on highly-textured regions; we use a linear combination of the shape and flow correlation distance to perform the behavior detection. The experimental results show that this method can reduce the false positive rate effectively.(4) Implement the whole system with Matlab, C++, Intel OpenCV in the Linux operating system. Our experiments on human actions, such as stand up or waving hands in examination room show reliable detection with few false positives.