Video Pedestrian Detection And People Counting

Video-based pedestrian detection and statistics technology,which can quickly detect the number of people and timely grasp the crowd information,plays an important role in security,business,and transportation.This paper focuses on this issue and the main research contains foreground extraction,shadow effect,crowd adhesion and occlusion method.The main work is in following aspects:First,foreground extraction.Both the characteristics of Gaussian Mixed background Model(GMM)and Mean background Model(MM)are analyzed.A foreground detection method combined GMM and MM is proposed to extract initial foreground.A following morphological process is designed to eliminate the noise in the initial foreground,thus a better initial foreground is obtained with a relatively complete pedestrian area.Second,for the problem of the shadow effect and slight adhesion between pedestrians in foreground,pedestrian segmentation and shadow removal methods based on the Bessel curve fitting is developed according to the multi peaks-valleys characteristics of the vertical projection cumulative curve for adhesion area,pedestrian,and shadow.Thus these two problems are solved,and more independent people are obtained.A block merging method is designed to eliminate the false alarm bodies.Third,for the foreground with high population density and serious obstacles,people statistics is achieved based on the mapping relationship between foreground feature and the number of people.A bidirectional self-calibration method based on fixed target is developed,and the perspective problem is solved without changing the shape of the image.The foreground database is constructed using test video,and the texture and geometric features are extracted and analyzed.A number estimate method based on geometric features and a number estimate method based on Adaboost has been developed to estimate the number of people with high-density occlusion.The comparison of the two algorithms shows that the geometric feature method is better with the single-frame estimation error is less than 35%,and the total number of errors is less than 13%.Fourth,A pedestrian number statistics system is built with MATLAB GUI by integrate the developed algorithms,pedestrian detection and number statistics for single-frame and continuous video are completed.The algorithm is proved to be effective,and the GUI system has the application prospect.