Reasearch On Human Detection And Behavior Recognition In Video Sequence

Intelligetice and automation of security video surveillance are widely studied in computer vision over the past several years, in which the human object detection, tracking and behavior recognition is the main research tasks of intelligent video surveillance. Intelligent surveillance system not only has important practical applications, but also play an important roles to artificial intelligence, pattern recognition, and other areas of computer vision. The tasks of object detection and recognition of human behavior are to detect human objects in the video, and identification and analysis of their behavior in real-time. Although after ten years of tireless research which from a large number of computer scientists, has formed a number of excellent pedestrian and behavior recognition algorithms, but practice shows that various human object detection, tracking and behavior recognition algorithms have more or less shortcomings, and would be limited by practical application scenarios, so propose a algorithm which can be applied to a variety of complex situations becomes the current urgent task. This paper studies the human object detection, tracking and behavior recognition which in key technical issues. The main contributions of this paper can be concluded as follow:1)A foreground segmentation algorithm is proposed based on three-dimensional ellipse codebook model. In this method could solve the problem which pedestrian’s shadow under strong light is difficult to remove by establish a 3D ellipse model for one pixel in RGB color space, and record the pixel’s frequency in the video sequence,and then assemble them together into a codebook. According to the value of the pixel is whether out of the 3D ellipse model or not, determining it is the foreground object. Experiments show that the improved method can effectively deal with the darker areas and noise in the video, and can be accurately segmented human body’s shadow in the strong light environment, for periodic-like dynamic background this algorithm can achieve better results.2) Crowd people segmentation algorithm in complex environments is proposed based on Bayesian formula. In this method solved the case of the occlusion of crowd pedestrian which are difficult to detect one pedestrian from each other. This algorithm uses the Bayesian formula to convert the problem that how to segmentation the crowd people into seeking maximum a posteriori probability. Using the assumption segmentation model of the pedestrian to match the foreground area, and then modify segmentation model for converging to the maximum posterior probability through multiple iterations. At last it would get the best crowd segmentation results. Experiments result show that the three videos with different degrees of overlapping pedestrians by using this algorithm can effectively and accurately detect the pedestrian whose occlusion is less half and the part of severely blocked pedestrians can also be identification.3) A human object tracking algorithm is proposed based on the color gradient direction histogram in the framework of the Kalman filter. This method could solve the problem that when tracking object is blocked would be leading to the failure. The algorithm uses a histogram model which is based on gradient direction to model pedestrian, this pedestrian modeling method could contain the body of texture information completely, so that pedestrians between two frames would be matched more accurately. Then combining with Kalman filter to predict the pedestrian position, so it could track the pedestrian which is blocked completely by predicting its future position at the follow frames and it can be effectively determined whether the pedestrian is the previous one when it reappears in the video. Experiments show that combining the color gradient direction histogram and the Kalman filtering method could accurately tracking object, and would effectively track the pedestrian which is completely blocked.4) Recognition of the human behavior in the video, according to the parts of moving body of the human which are changing dramatically in time and space, at first extracting the spatial-temporal interesting points which are scale invariance, and then establishing the feature descriptor by the information of the neighbor of the spatial-temporal interesting points feature points. The feature descriptors which extracting from videos which have different behaviors would be had different distribution in the space, so building Gaussian mixture models of descriptors by the EM algorithm according to this characteristics. And then get the feature vector by statisticing descriptors’frequency of distribution in the Gaussian mixture models. A likelihood is computed by histogram intersection method to determine the behavior. Experiments show that this algorithm has a high recognition rate for simple abnormal behavior.