Extraction And Recognition Of Human Skeleton And 3D Modeling Based On Video Sequence

Recently as the basis of recognition in machine vision field, Real-time object tracking becoming more and more important in the video analysis and processing field. This research has various applications in human animation, computer game, virtual reality and augmented reality, human-computer interaction, video surveillance, ports video analysis, computer-aided clinical diagnosis, and so on. So studying the extraction and recognition of human skeleton and 3D modeling based on video sequence images is very important.There are three basic components of tracking human movement from sequence images: (1) Extracting human motion region from the complex background; (2) Human body tracking and demarcate; (3) Recognition and understanding of human behavior. Human body tracking and demarcate is the key of human movement analysis. It is also the basis of recognition of human movement. The system accomplished in this paper is based on two cameras. The input is two sequence images. Firstly, the background difference method, combined with histogram for threshold segmentation is used to detect the contours of the human body. Then human morphology is carried out to get better result. Secondly, according to the contours of human body, extract the key points of human body. The vectorization of body contour and apar are used to divide body into several sub-regions. Skeleton algorithm is also used to filter apar seeds and get the key points of body. At last, reconstruct the 3D human motion skeleton sequence. Firstly, establish the link between images, followed by the recovery of structure and movement, then camera self-calibration, finally 3D reconstruction is finished.In this thesis, mark-less technology is applied. And there are no markers or sensors on the body. Because the first image does not need artificial calibration, customer interference is reduced, and intelligence is high. Experiments show that it can accurately identify simple movement, and reconstruct to 3D profile. It also can resist some noise interference, meet the needs of real-time. But usually there are some key points overlapped, which make it uncertain. Even if use multiple cameras, overlapped can't be solved completely. However, this thesis is meaningful to human body motion capture.