Human Body Detection And Posture Recognition Based On Low Quality Videos

Human tracking and posture detection has been one of the most popular research direction in the field of intelligent monitoring. It aims to do people foreground extraction, characteristics analysis and human posture recognition to the video sequence recorded by the camera with the method in Computer Vision field. The content of this paper which is based on the video sequence recorded by low quality fish-eye lens, starts researching on recognizing and tracking dynamic people and static posture detection. It has many potential applications in intelligent monitoring, human-computer interaction, IA.Because it is affected by low quality lens, rare samples, big noise from background in video samples and indeterminacy direction of motion, after consulting a large number of domestic and foreign periodicals and literatures, the author uses gait detection to track and recognize dynamic people, and us-es ellipse detection and color projection vector to detect static people posture with improving Gaussian Model method. Some novel algorithms are pro-posed in the research progress, which leads to ideal experimental results.(1) Discuss the content of background model in video pretreatment, and analyze the classical Gaussian Mixture Model and improving Gaussian Mix-ture Model modified by Xiaodong Cai, this paper adds an innovative param-eter restoration algorithm to Gaussian Mixture Model according to the needs of the video samples. It successfully modifies Gaussian Mixture Model by bringing in a feedback adjustment link, and making the model have the abil-ity to recognize multi-modal background and static target.(2) Discuss the content of shadow detection in video pretreatment, after the comparison with a large number of shadow detection methods, this paper find the most suitable adjustment characteristic by sample statistical experi-ment, which is capable of filtering over 90% shadow area.(3) Discuss human gait detection method, and divide it into 3 parts to help each other recognition results. This paper uses human GEI as basic fea-ture, then uses PHOG to do further processing and choose PCA and LDA di-mension reduction method to get the contour feature; it extracts SVB ribbon frieze pattern and both pace difference pattern as foundation, uses dynamic time lining-up to match, and get motion feature; then it will conjoin the space-time feature to do gait recognition. Experimental results show that this algorithm will recognize people ID accurately.(4) Discuss the static feature contained by human foreground contour. This paper uses ellipse detection to get space-time feature in foreground based on binary contour secondary moment, and designs a novel feature based on the projection length of clothes color, which will help the former to recognize static posture. After analyzing a large number of experiment sam-ples, a most suitable classification process and threshold are found. Experi-mental results show that this method can recognize 4 human postures accu-rately which includes standing, lying, sitting on the ground and sitting on a chair. It has low complexity of time and can be calculated in real-time.