Research On People Counting In Real-time Video Surveillance

With the development of society and economy, video surveillance has been applied in many aspects in our everyday life. It is a trend to use the computer vision to improve the automation of video surveillance system and reduce human intervention in the future. Using the video surveillance system as a platform, it is of extremely significance to get the accurate number of the pedestrians who are detected and tracked in the video sequences. Recently, the study of people counting by video surveillance has become a hotspot with great difficulties in computer vision area.This thesis mainly researched the image processing algorithm related to people counting by video surveillance. Through analyzing the existing moving targets detection and tracking algorithm, a people counting system was designed in the thesis. The people counting algorithm was divided into three parts which involved moving targets detection, moving targets tracking and trajectory analysis.Moving targets detection was the basis of the people counting system in the thesis. Since the application environments of the video surveillance comprised stationary background, we exploited background subtraction approach to extract the moving targets. This paper focused on the selective background update model and Gaussian mixture model, and compared these two models. The selective background update model was improved and a robust, real-time background estimation method was proposed. After having got the estimated background, the colored background subtraction approach was used to extract the targets. When it came to Post-processing for the foreground image: first of all, the algorithm made use of morphological filters to remove noises, then eliminated the shadow of moving targets using the brightness and the normalized colored information of the pixel, and finally used the motion history image to detect the targets more completely.In the aspect of moving targets tracking, the thesis studied on tracking algorithm based on regional characteristics, which made use of multiple features of local area to match the target. To achieve stable tracking of the targets, the algorithm must overcome the impact of block, so this thesis divided the block effect into static and dynamic block and conducted the research from this two aspects, then focused on analyzing pedestrian merger, separation and other key issues, which influenced the accurate people counting very much, and finally validated the effectiveness of the algorithm through experiments.After the realization of pedestrians tracking, this thesis analyzed the trajectory of the target and presented an effective people counting method which could count pedestrians accurately and overcome the impact of complex motion of the targets.Real-time performance and practicality was the goal of this thesis, a number of key technologies in people counting in video surveillance were studied and a robust method with low complexity was proposed, which achieved good results in the experiment.