Monocular Depth Estimation And Depth Completion Based On Convolutional Neural Network

Scene depth estimation is an important topic in the field of computer vision.For many computer vision tasks,adding additional depth information to the 2D image can simplify the task and effectively improve the performance of the algorithm,such as object detection,video surveillance,and semantic segmentation.This paper focuses on obtaining dense depth information from single-view color image or sparse depth map.The accuracy of monocular depth estimation algorithm based machine learning is low due to poor model fitting ability.In outdoor large scenes,depth information is generally generated by a 3D point cloud projection captured by a lidar into a color map.Due to the limited number of point clouds scanned during the unit period,the resulting depth map is very sparse.Aiming at these two problems,monocular depth estimation and depth completion based on convolutional neural networks are proposed.The main work content and innovative achievements of this paper are as follows:(1)A dilated residual convolutional neural network based on uncertainty is proposed to improve the prediction depth accuracy of traditional machine learning methods in monocular depth estimation.The combination of dilated convolution and skip connections not only improves the overall accuracy of the predicted depth,but also performs better in edge details.In addition,a method of uncertainty learning is proposed to solve the problem that the current convolutional neural network for monocular depth estimation has strong expression ability but cannot evaluate the reliability of the output.By modeling the uncertainty,the method can predict the confidence of the estimated depth and improve the prediction accuracy of the model.(2)A depth completion algorithm based on semi-supervised learning is proposed to solve the problem that the depth map generated by the lidar projection from the point cloud to the color image space is very sparse.The method uses a color image containing rich scene information to guide the network to estimate the dense depth map,and overcomes the input sparse depth map,which can only express information of a small part of the region,and improve the prediction accuracy of the model.A post-fusion method of color image and depth map is proposed to solve the problem that direct fusion of two kinds of images is useless for performance improvement due to differences in expression content,expression range and expression ability.(3)A monocular dense reconstruction system based on CNN estimation depth is proposed to improve the performance of traditional monocular reconstruction methods in low-texture regions or pure rotation,and a more realistic reconstruction effect is obtained.