Research Of Pedestrian Detection Algorithm Based On On-board Vision System

With the increasing of the urban population and the number of vehicles,safe driving has become a social problem.Since pedestrians are the main victim group of traffic accident,it has great practical significance to study pedestrian in the road environment detection algorithm.And pedestrian detection technology which is one of the critical technologies in safe driving has become the leading direction in the field of intelligent transportation and safe driving.In this research,we mainly detect the pedestrian through the images obtained by vehicle-mounted camera.Different with the pedestrian detection of other fields,pedestrian detection of vehicle driving environment has a high demand on the detection rate and real-time,so that it has more challenges: there are too many small-sized pedestrian in the vehicle image to identify;the background of urban street is too complex to understand easily;the limited computing resources in the vehicle environment bring greater pressure to detection algorithm.In this paper,were search and improve pedestrian detection algorithm through three stages including the generation of candidate windows,pedestrian feature extraction and feature classification.We have improved the exhaustive window scanning approach: at first,we generate candidate windows using the image gray information,and then pretreatment algorithm is proposed based on the prior knowledge of the scene.In the feature selection stage,we combine a variety of feature information together to describe the pedestrian in order to better identify pedestrian with low resolution.In the calculation of multi-scale feature,we adopt a fast feature pyramid algorithm to speed up feature calculation.Aiming at the problem of large number of disturbing objects in environment background,we propose a feature extraction method based on human template filtering: the contour information of the human body is used as a priori knowledge to generate rectangular filter banks for enhancing the expression of feature.At the same time,a fusion algorithm of rectangle filter template with other non-priori knowledge is designed and the optimal filter template banks are obtained by training several times.In the stage of feature classification,we use Ada Boost to integrate multiple weak classifiers and soft cascade method to optimize cascade classifiers,and adopt multi-round iterative method to accelerate the training of the classifier.We have compared the results of the algorithm proposed with other algorithms on different public datasets.The results show that the algorithm proposed not only achieves better detection accuracy,but also have a strong ability to adapt to the different scenes.At the same time,the proposed algorithm achieves good results on the data sets obtained in the actual street scene,which indicates that the proposed algorithm has certain practical value.