Technology For People Moving Behavior Modeling

As two important fields of computer vision research, human motion detection and behavior recognition which have great application prospects and economic benefits in intelligent video surveillance, human-computer interaction, content-based video compression and image retrieval are concerned about by a growing number of scholars. In this paper we have systematically studied how to detect moving human object in the video and recognize human moving behavior, such as walk, run, bend jump and so on. According to the practical application of human motion detection and behavior recognition, this study includes the following:1、Motion detection and extraction in complex scenes. According to different application scenarios, two detection methods are proposed, one based on the color component statistics and the other one based on local texture. For solving the problems that the popular background modeling algorithm of codebook and mixture of Gaussian could not accurately describe the distribution of the dynamic background, the process of background modeling of the two methods is too complex, and they are low real-time when deal with high-resolution pictures, a fast and simple background modeling algorithm based on the color component statistics is proposed. By the analysis of the RGB color components of the dynamic background pixels, we noticed that the difference between any two of the three components fluctuates in a very small range. So based on the fact, we proposed a background modeling method based on the RGB color component statistics. For the single-pixel distribution analysis and statistical methods on the timeline cannot obtain satisfactory results in the face of the global illumination changes and shadow interference cases, we propose a pixel-based background subtraction method based on local texture pattern. By improving the texture descriptors of scale invariant local ternary patterns, it has a more robust description of the local texture and by extending the spatial domain texture pattern to the temporal domain, it has better discrimination against the foreground.2、Body contour shape description and shape similarity measure. We propose a body contour description method based on the direction function of the arc length and a human shape matching method based on geodesic distance. Firstly, the methods of contours curve representation such as point set description, direction functions and curvature functions are introduced. Then we use the Fourier descriptors based on the points sets to smooth the contour and eliminate small noise on the contour. And then a equal length segment method to fit the body contour is used to compute the direction function about the curve length. Lastly, we propose a human contour similarity measure method based on geodesic distance.3、Action recognition based on the human silhouette. Based on the body shape description and similarity measure we propose three methods to achieve the recognition of human action which respectively are key-frame-based behavior recognition method, template matching based method and hidden Markov model based method. Key-frame-based action recognition method is achieved by extracting, training and matching the key frames. This method is simple to implement for a simple cyclical behavior and with the advantage of high real-time, however the recognition results is dependent on the accuracy of the key frame extraction. The template matching method is achieved by clustering contours to train the body posture template, and by dynamic time warping method to match the posture sequence. For more complex action, the method based on hidden Markov model can better describe changes in the timing of the action and has a high recognition rate.4、Human action recognition under dynamic scenes. When the camera is moving or the background is badly dynamic, extracting the accurate human silhouette is a difficult task. We proposed action recognition method based on the integration of two methods. The one is the space-time interest point which is used to describe the local feature of action and the other one is histogram of oriented gradient which is used to describe the feature of human global feature. By the test on KTH database, a higher recognition rate is achieved.