Human Contour Recognition Technology Based On Depth Map

Visual analysis of human motion is a hot topic in recent years, which has attracted much attention in the field of machine vision. It has broad application prospects and potential economic value in areas such as in human-computer interaction, video conferencing, medical diagnosis, game animation, virtual reality, and security of community, supermarkets, banks, monitoring image storage, which has aroused great interest of researchers and related enterprises.Because body contour recognition is an important part of human motion, the accuracy of recognition will directly affect success of visual analysis of human motion, so it has important theoretical value and broad application scope. The traditional human detection technology is mainly based on RGB color image by visible light camera, robustness is not high. The main reason is that it would be easily affected by illumination changes, the influence of complex environment and object occlusion. This paper designed a multiple moving targets detection algorithm in complex dynamic background. This body contour detection algorithm is proposed in this paper can be used as a preliminary work of human gesture detection, tracking and behavior recognition. Its research has a good theoretical significance and engineering application value. In this paper, we also transform the two-dimensional coordinate information of the human body into three-dimensional space coordinate which can provide further support for the recognition of human behavior. Experiments show that this system can detect the human body in the laboratory real-timely which is a complex environment, and the performance of system is not affected by the factors such as light.The main research of this paper is as follows:1) Because the initial depth map acquired from Kinect has a large number of "black hole", noise and other defects, we studied fast marching repair algorithm FMM, corrosion, expansion and other commonly used image processing algorithms. According to the value of pix in depth map equals distance from object to the camera body, we divide the map based on that distance and detect contour to acquire different depth contour collection. Then we finish preprocessing of different size of template matching.2) Based on the study of the human body detection, we studied on the distance between contours and the Hu moment feature matching technology. Through the head contour template and real-time contour matching selection, we complete the body contour localization.3) In order to ensure the real-time detection of the contour of the human body, we studied the coarse-grained parallel genetic algorithm. We convert the contour similarity into the fitness function of genetic algorithm. We initialize a number of agents and find the best matching position by individual evolution.4) We studied the conversion between Kinect coordinate system and the image coordinate system. We convert the coordinate of human body contour information into 3D space coordinate. We think it can provide further support for the recognition of human behavior.