Man-Shaped Object Detection In Intelligent Video Surveillance

Moving object detection is one of the core knowledge in modern video surveillance system, and it has very important theoretical significance and application value to the follow-up study on intelligent video surveillance. The man-shaped object detection and face detection have been of widespread concern and attention in the fields of pattern recognition and computer vision. This thesis studies man-shaped object detection in intelligent visual surveillance, and the main work completed in is as follows:(1) For the deficiencies of traditional moving object detection methods in intelligent video surveillance, a fast and efficient method is presented. First of all, because the background is not fixed in the moving object recognition system, it will reconstruct the background image with consecutive frames subtraction and separate the background and moving target of the current frame image with background subtraction. Then, the error rate of this module is reduced through filtering the discontinuous moving targets. At the same time, self-adaptive updating of background is adopted to update the background at regular intervals, in order to achieve the desirable effects of partition.(2) The thesis adopts a weight update method with adaptive threshold based on AdaBoost algorithm and Haar-like feathers, optimizes the structure of the classifier to achieve a cascade classifier training method and reduces the algorithm complexity. The background module is achieved through the improved moving object detection which can help exclude a large number of objects in the background and get moving man-shaped objects in video sequences. It helps complete a man-shaped object detection system ultimately. Experimental results show that using Haar-like rectangular feature can overcome the shortcomings of the traditional detection methods, which can not accurately extract the features in noisy conditions. The method can get good detection results in the atypical weather scene, single or multi-object in snow days and other complex scenarios.(3) The thesis presents an AdaBoost algorithm based on skin-color space pretreatment which uses skin-color model to extract the skin-color area at first and establishes the Gaussian model of skin-color in the YCbCr color space. After processing noise of the image, it will get face area after the preliminary examination through skin-color feature analysis at last. In the AdaBoost detection phase, a more accurate face location can be got after the OR operation between the face area obtained and the previous face area obtained based on skin-color detection. The experimental results show that this algorithm can be used for single face, multi-face detection in the image and single-multi-angle,multi-face detection in the real-time video sequences. And it can accurately locate human faces in video sequences and can obtain better detection rate and low error rate.