Automatic People Counting For Video Surveillance

As a typical application in the computer vision field, video people counting is of great value in many aspects such as people counting business data statistical analysis, intelligent transportation and public safety. It can’t meet the current requirements by relying on the artificial pedestrian statistical methods. People counting with high efficiency and high accuracy can be achieved by use of the computing power of computers and the advance of techniques in computer vision. Many pedestrian detection algorithms have been proposed recently. However, in real situations, it is still hard to achieve high accuracy pedestrian detection since it is vulnerable to occlusion, gestures and environment factors.This thesis studies the pedestrian detection based on image features. It focuses on the pedestrian detection in static images and video to improve its accuracy. The main research work and innovation of this paper is mainly reflected in the following aspects.1. A pedestrian detection method based on multiple decisions is proposed. It is an improvement of Algorithm C4on the detection accuracy. Its basic idea is:A classifier is firstly used to classify a kind of feature. Then whether the classification value belongs to the suspicious interval or not is determined. B classifier is further employed to classify another feature if yes. The value of the result is determined by two decisions together. Experimental results show that the method of multiple decisions can improve the detection accuracy while it hardly affects the detection speed of Algorithm C4.2. A pedestrian detection method is presented based on the profile calibration. Three different calibration ways (manual calibration, random calibration and average calibration) are studied by the number of calibration points and combination of features. Experimental results show that the best one is the manual calibration by applying combining features (LBP and BHOG).3. A people counting based on dynamic video is achieved by using Gaussian Mixture Model and Tracking approach, respectively. The foreground region is firstly extracted through Gaussian mixture model. Then a more accurate foreground region is obtained by processing the regional connectivity. Finally, the static pedestrian detection is carried on this region to achieve the people counting. The Gaussian mixture model method performs the best in the video with fast moving people. It does not perform well when the pedestrian movement is very slow or still. It needs to further research on the background modeling.4. A pedestrian detection and counting system is developed. It is divided into the feature extraction module, the model training module, the detection module and GMM module. It can not only display the detected pedestrian, but also output the pedestrian position.