Road Obstacle Detection And Size Estimation Based On Monocular Depth Estimation

Environmental perception is the foundation of the intelligence of engineering vehicles.Detecting obstacles on the road and estimating their size can help avoid potential hazards and reduce the risk of accidents.The 3D object detection methods based on Lidar are becoming more and more mature.However,the high price of Lidar limits the application of related methods in industrial production and life.To solve this problem,the thesis uses a monocular camera to replace the expensive Lidar.The selfsupervised monocular depth estimation model predicts the depth map and converts the depth map into the pseudo point cloud for road obstacles detection and size estimation.The thesis proposes improvement measures for the problems existing in the process of depth estimation and pseudo-point cloud data conversion.The main contributions of the thesis are as follows:Firstly,to solve the problem that the self-supervised monocular depth estimation method lacks global information for supervision,this thesis proposes an automatic pseudo-label generation method based on feature point matching and geometric view relationship.This method can generate pseudo-sampling grid labels and pseudo-pose labels at the same time.The pseudo-sampling grid labels can directly calculate the loss with the sampling grid.The thesis further designs a two-stream pose estimation network,which can use the pseudo-pose label and the pose generated by the network together for training,so that the model can learn the information based on network and geometry.The automatic pseudo-label generation method can learn more global structured information and provide global supervision information for the model.Secondly,to solve the problem of the position offset of the sampling coordinates in the low-texture area,this thesis proposes a depth regularization method guided by the sampling grid.The thesis takes advantage of the ideally uniform distribution of the sampling grid in the low-texture area.Wrong depth estimation will break such ideal distribution.So this thesis uses the sampling grid to guide the regularization of the depth map of the area.The depth regularization method can effectively alleviate the problem of the position offset of the sampling coordinates in the low-texture area.Thirdly,the depth map is converted into pseudo-point cloud data,and on this basis,an adaptive ground plane extraction method based on the V-disparity map is proposed to assist the calculation of height features.The thesis generates pseudo-point cloud data from dense depth maps,which reduces the cost of sensing equipment and obtains more point cloud information.At the same time,the thesis proposes a ground plane extraction method based on the V-disparity map,which can adaptively extract the range of the ground plane area.The method solves the problem that the fixed-range ground plane area determined by experience is not applicable,filters noise data,and improves the accuracy of the ground plane fitting.