Research On Road Pedestrian Detection Based On Fusion Of Lidar And Visual Information

The stable and reliable perception ability is the basis of the safe driving of intelligent vehicles,and the target detection is one of the key technologies of unmanned driving in t environment perception.At present,the target detection based on vision has a good performance,but due to the defects of the visual sensor itself,the detection effect is easily affected by the external light and weather conditions.At the same time,due to the limited field of vision,the existence of detection blind area,the lack of means to obtain global information and other shortcomings,it can not accurately provide the target location information.The LIDAR can accurately obtain the three-dimensional information of the target,and has high reliability and accuracy,and is less affected by the light and weather conditions,so the two sensors have strong complementary.Therefore,the establishment of a pedestrian detection system based on LIDAR and visual information fusion has important research significance and engineering value for driverless vehicles.The specific work is as follows:(1)In order to solve the data dependence problem of the existing target detection model,and improve the adaptability and robustness of the model under different environmental conditions,this paper proposes a data enhancement algorithm based on style transfers,and uses this algorithm to enhance the training of yolov3 target detection algorithm.This method can change the texture,contrast and color of the image.Compared with the traditional enhancement technology,it can not only improve the diversity of limited data,but also improve the generalization ability and robustness of the target model more effectively.(2)Aiming at the problems of poor detection speed,accuracy and high equipment cost of current LIDAR point cloud recognition network,a 3D point clouds recognition network based on yolov3 point is proposed.A fast and convenient reprocessing module is added to the network.Compared with the traditional point cloud recognition method,it avoids the time spent on the extraction of the region of interest and the removal of the ground point cloud,and can directly identify the input point cloud data globally.Finally,on the premise of ensuring better detection accuracy,the detection speed is improved.(3)Aiming at the environment perception system of driverless vehicle,this paper studies a pedestrian target detection scheme based on multi-sensor information fusion.In the process of target detection,data association is used to establish the relationship between the multi-sensor target detection results.Combined with the historical detection results of the target,the target classification and confidence are finally determined according to the method of DSm T evidence theory.Then,relying on Guizhou University driverless car platform,the overall layout of Velodyne 16 lines LIDAR and vision sensor is carried out.The experimental results show that the proposed fusion method is reasonable and effective.