| Human motion analysis based on computer vision has been a significant research direction in machine vision and image processing during the past few years. Human motion analysis system usually involves image pre-processing, object detection, tracking and recognition, comprehension and description of movement behavior. The main purpose of development of computer vision is to make the communication between computers and human smooth. Meanwhile, computer vision should make computers understand the language, gestures and other actions of human. Finally, computers should be able to accomplish the task beyond human brain. Human motion analysis system has broad application prospects and economic values in virtual reality, smart surveillance, human-computer interaction, video retrieval, medical diagnostics and other fields. However, due to the complexity and particularity of human movement, the research is still in its infancy. For example, in the process of movement tracking, experimenters need to mark the initial point of initial frame manually; In terms of movement recognition, the machine learning methods which are suitable for human movement analysis are defective.Combined with the current research status and characteristics of human motion analysis, this paper studies human motion analysis in fixed background on monocular vision and carries out the innovative work in target detection, human action recognition and target tracking. The main work of this paper includes:1. In respect of moving target detection, the paper summaries the current moving target detection techniques, shows the experimental results of partly algorithms, and focuses on the analysis of background subtraction method under monocular camera. Primarily, the paper obtains a preliminary outline of human body through the background subtraction method, and removes noise through mathematical morphology. In order to obtain smooth target contours, the paper executes the algorithm of judging the size of image connected domain, sets a specific threshold, and removes noise block connected domains which are smaller than the threshold.2. In respect of human action recognition, the paper selects10body motion characteristics, including aspect ratio of minimum outside rectangle, rectangle degree, circular degree,7Hu invariant moments. The selection criteria of human motion characteristics are strong in anti-noise immunity and high distinction. Three types of human motion images are classified and identified by SVM. The recognition accuracy improves significantly through the cross-validation and parameter optimization.3. In respect of moving target tracking, the paper shows specific tracking algorithm based on Mean Shift and Kalman filter, and has drawn a trajectory tracking line of walking motion. In order to evaluate the effect of tracking algorithm objectively, the paper has calculated the Euclidean distance between the center of the actual body and the points of Mean Shift tracking algorithm, and then executes division operation of Euclidean distance and diagonal pixels distance of the original image.Finally, human motion analysis system is realized by the programming under Visual C++6.0. Experimental results showed that the algorithm proposed by the author has high accuracy in motion recognition. The effect of human tracking algorithm is significantly improved and results better than the traditional Mean Shift algorithm. |
