| In automatic driving,surveying,security monitoring and other three-dimensional visual fields,the environmental perception module,which obtains color and depth information from sensors,is the prerequisite for subsequent tasks such as scene understanding,decision planning and motion control.However,as the core sensor,Li DAR with more channels can capture more details but is expensive,and cannot obtain color information to complete environmental perception alone.This paper focuses on improving the perception precision and reducing the cost,two methods based on convolutional neural network are proposed to increase the density of Li DAR point cloud,and the latter can be combined with the camera to obtain color information,integrate them,and make improvements.The main research work of this paper is as follows:A Li DAR super-resolution method based on deep learning is proposed.This method can simulate advanced Li DAR only by using one with fewer channels.The improved convolutional neural network is trained using the data generated by the virtual world simulator.A low-resolution two-dimensional range image with sparse depth information is projected by a three-dimensional Li DAR point cloud,which is input into a pre-trained network to output a high-resolution range image.This method can be used in the data preprocessing part of the environmental perception scheme in this paper,and its feasibility is verified by experiments on real environmental dataset.A depth completion method based on the fusion of Li DAR and monocular RGB camera is proposed.Aided by a cheaper camera,this method can obtain color and denser depth information.By using the global network to grasp the overall contour,and the local network to distinguish the detail features,the feature extraction results of the two branches are normalized with the confidence weight to obtain the completed depth image.The feasibility of this method is crossverified by experiments on outdoor scene KITTI dataset and indoor scene NYUDv2 dataset.The environmental perception scheme integrating these two methods also considers the failure of joint calibration relationship caused by sensor motion deviation.The front-end visual odometry and back-end optimization in SLAM are constructed to solve the real-time pose relationship between sensors to correct the deviation,thus realizing the optimization of the scheme.The ablation experiment verifies the effectiveness of each method for improvement.The complete scheme performs best in KITTI depth completion test,its RMSE is 722.43 mm,MAE is 205.84 mm,and the runtime is 0.1s.Compared with other advanced schemes,the proposed environmental perception scheme balances precision,real-time performance and cost better. |
