Pedestrian Counting In Crowded Scenes

Real-time crowd counting is a research topic with theoretical significance and many potential applications,such as surveillance safety,crowded flow control in subway stations.There exist a number of difficulties in crowded scene pedestrian counting.First,in crowded scenes,the occlusion between people is serious.Second,the resolution of video in surveillance camera is relatively low,detailed information is lost.Third,because the angle of the camera is not frontal,pedestrians will be distorted.Forth,under different scenes,the background,light condition and appearance of pedestrians may be very different.We analyze different methods for pedestrian counting in both images and videos.To satisfy the demand of real time in real-word applications,we propose a novel pedestrian counting method.This method uses human detection and human tracking to count pedestrian.The proposed approach achieves high accuracy and is easy to deploy.We assume that the cameras are mounted at the front of pedestrians' head,the angle is between 30 and 60 degree.This configuration is in accordance with the settings of cameras in most markets and public places.To acquire an appropriate resolution,we need to select a region with reasonable distance,ignoring both the far distant region and close region.This paper mainly focuses on the following works:1)Human detection.Human detection is used to detect pedestrians in every frame.To achieve the goal of real time,we adopt the harr-like features to train a cascaded Adaboost classifier.2)Human tracking.Human tracking is used to connect pedestrians in different frames.According to the characteristics of pedestrian counting,we set up some prior knowledge,so we improve the speed of human tracking.3)Experiments.We recorded videos in the subway station and evaluated the proposed algorithm on this data set.The experiments show that our algorithm achieves a high accuracy and it can run in real time.We also compare our algorithm with an algorithm using low level features and the regression method to count pedestrians.Experiments show our algorithm achieves a higher accuracy.