| The development of artificial intelligence has promoted the development of self-driving cars,which not only bring convenience to people’s lives,but also reduce traffic accidents.Self-driving technology is an advanced stage of intelligent vehicles,the environment perception layer is one of the key technologies of intelligent vehicles,so many Research Institutions and companies fouce on the environment perception layer.Based on low-speed intelligent vehicle research platform,using Lidar and camera to collecte the obstacle information on the track,the researcher has completed the point cloud clustering processing and obstacle recognition respectively.At the same time,using Lidar and visual information fusion recognition algorithm to complete the fusion recognition of track obstacles,the researcher has obtained the information of environment perception layer with high reliability and stability.And then,the main research contents of this paper are as follows:1.Spatio-temporal Synchronization Processing of Lidar and CameraUsing Zhang Zhengyou’s Checkerboard Grid Method to calibrate the camera to obtain the camera internal reference,the joint calibration of Lidar and camera was completed by using MATLAB toolbox,and the frequency synchronization of Lidar and camera data was completed by using the Lidar backward compatibility method,so as to achieve the synchronization of Lidar and camera in space and time.2.Point Cloud Clustering Based on LIDAR TrackThe Lidar is used to collect the information on the track,obtain the point cloud data,establish the region of interest based on the track area,complete the ground point cloud segmentation using the RANSAC algorithm,and complete the clustering of the point cloud on the ground using the DBSCAN clustering algorithm that automatically calculates the initial radius and the minimum neighborhood Minpts,so as to achieve an effective clustering of obstacles.3.Image Recognition Based on Convolutional Neural NetworkThe input images are subjected to Mosaic data enhancement,Backbone network convolution and Neck network pooling using convolutional neural network YOLO-v5 s algorithm to output multi-feature images.The model is trained using manually collected data and KITTI dataset,and the trained model is tested against obstacles on the track.The test results show that the accuracy and recall curves of the model fluctuate smoothly,and the mean accuracy MAP value reaches more than 80%,and the model can meet the task of accurate target recognition.4.Identification and Experiment of LIDAR and Vision Information FusionBased on the principle of complementary advantages of Lidar and camera,the obstacle recognition method of Lidar and vision fusion is proposed.To verify the feasibility and reliability of the method,information fusion recognition test and single sensor and information fusion recognition comparison test are conducted in the low-speed intelligent vehicle respectively.The test statistics show that the recognition rate of camera is 93.63 %,the correct rate of Lidar clustering is 90.34%,and the fusion is 95.26%,which verifies the feasibility of information fusion. |
