Research On Detection Method Of Human Abnormal Behavior Based On RGB-D Video

Since 2009, with the use of Kinect and other depth sensors, more and more scholars try to help achieve more accurate target detection, recognition, tracking and video analysis by using RGB-D depth images in the field of computer vision research. The RGB-D depth images contain three-dimensional information and they are robust in situations when light,shadow and conflict background change, which makes the extraction of human skeleton features easier than using common RGB images. Based on that, the human abnormal behavior detection using RGB-D images have become a research focus.The main work and new achievements of this paper are as follows:1.The principle and process of the Kinect to obtain the depth images are introduced.And representation of RGB-D depth images using the color point cloud method and pseudo color method are studied and realized.2.The principle of the human skeleton feature extraction algorithm based on RGB-D depth image are introduced. Based on this, we propose the method of the joint angle feature of human body posture. Experiments have been done to verify the effectiveness of the human body skeleton feature extraction in the different perspective, different distance,human falls, and more human targets.3.The definition and characteristics of abnormal behavior are studied. According to the characteristics of high disorder, an algorithm is proposed to detect the human abnormal behavior based on human skeleton feature information entropy. The experiments show that the algorithm has a robust result whose recall is 92% and precision is 95.83%, and accuracy is 94%. Based on this, a new method is proposed using the human spatial position information, human body motion velocity and human joint angle information, which can be used in intrusion detection, detection, retrograde detection, fall detection.etc.4.On the basis of the above research and experiments, the detection system of human abnormal behavior in RGB-D videos is designed, which includes the RGB-D video capture module, the human skeleton feature extraction module, the human abnormal behavior detection module and the video output display module.Finally, the summary and prospect are given.