Research On Key Technologies Of Pedestrian Detection Based On Binocular Stereo Vision

With the rapid development of economy and the large increase of population,many roads are used by pedestrians,non motorized vehicles and motor vehicles.In various traffic accidents,the casualties of pedestrians,non motorists and other disadvantaged traffic participants are more serious than drivers and occupants.Pedestrian detection has become a common research hotspot both at home and abroad.As the core component of the intelligent driving auxiliary system and the pilotless system,it is an effective and active safety auxiliary system for the protection of pedestrians and other road users,and has the practical significance of reducing the casualties of road traffic accidents.In this paper,the pedestrian detection in the traffic scene based on binocular stereo vision is proposed on the basis of the theory of pedestrian detection method proposed by predecessors,using parallax information to generate scene expression based on bar like pixels(stixels)can effectively reduce the candidate region of pedestrian detection and avoid the calculation of the depth map of the pixel level(pixels).In the preprocessing stage,the algorithm is significantly reduced.Each frame of the preprocessing phase is fused with Multi-HOG feature +LUV color space feature + hik-SVM feature fusion target detection strategy,based on the improvement of the traditional HOG features,the color features are fused,which improves the real-time performance of the algorithm and ensures the accuracy of the detection,so as to achieve real-time,accurate and highly robust pedestrian detection in traffic scenes.The main contents of this paper include the following aspects:(1)The advantages of parallax and depth information can be obtained by using binocular stereo vision.Combining the common characteristics of pedestrians in the traffic scene,the Stixel-world scene expression is used to avoid the pixel level calculation,and the complexity of the algorithm is reduced.(2)The dynamic programming algorithm is used to optimize the rationality of Stixels,and the region of interest is extracted effectively.Compared with the traditional algorithm,the candidate region of pedestrian detection is reduced greatly,and the detection area can be reduced to 73% in some scenes.(3)In order to ensure the accuracy of the detection,the optimized selection of the block scale to form Multi-HOG features and the fusion of LUV color channel features.The experimental results show that the error rate of the overall pedestrian target is less than 10% under the more complex environment,the detection rate is below 15%,and the accuracy of the detection is greatly improved.(4)The hik-SVM classifier significantly improves the real-time performance of the system in a large number of samples.The detection of TUD and ETH databases is shortened by 76.9% and 79.8% respectively,which basically meet the real-time requirements of the vehicle pedestrian detection system.