Research On Depth Completion Algorithms Based On Laser Radar And Camera Fusion

Scene depth information has important applications in the field of 3D vision,such as automatic driving and scene modeling.High-precision,high-density depth information can help the algorithms achieve better results.However,limited by physical equipment,dense depth information cannot be measured directly.The research in this paper mainly focuses on the depth completion technology of lidar and camera fusion,which predicts dense depth map through RGB images and sparse depth maps.Our main research work is as follows:A depth completion algorithm based on 3D point cloud features and guided filter fusion is proposed.By using the feature extraction method of 3D point cloud based on Pointnet and combined with the guide filtering algorithm,a 2D and 3D feature fusion method based on guided filtering is proposed.Then,the spatial propagation network based on affinity matrix and the local affinity are used to optimize the depth map,and a depth completion network based on 3D point cloud features and guided filter fusion is proposed.Finally,experiments are carried out on KITTI dataset and NYU dataset to analyze the effect of each module,and compared with the current mainstream methods to prove the effectiveness of this method.A self supervised graph propagation depth completion algorithm based on ground completion is proposed.Firstly,according to the simplicity of ground feature,a ground completion algorithm based on RANSAC is proposed to complete the depth value of ground area.At the same time,a self-supervised module based on ground completion algorithm is proposed.Then a depth completion algorithm based on graph propagation fusion is proposed.Finally,experiments are carried out on KITTI and VKITTI dataset,and the results show that the method is effective.The depth completion algorithm is applied to the loop closure system.By introducing the completed depth map as the feature,the accuracy of loop closure in low visibility environment is enhanced.In order to achieve real-time detection,the existing depth completion algorithm is optimized in speed.Finally,experiments are conduct on the loop closure system based on depth completion algorithm and the experimental results are analyzed to prove the effectiveness of this method.