| Depth information is one of the important elements of human perception of the world,through the binocular parallax and the accumulation of prior knowledge,human beings can easily obtain three-dimensional depth information.Obtaining accurate and dense depth information under the condition of only monocular images or sparse depth maps is still a very challenging problem in the field of computer vision.In tasks that require understanding and decision-making about a scene,it is important to obtain dense depth information at the pixel level in the scene.Due to the limitations of depth perception technology in terms of both hardware and algorithms,existing common instruments for obtaining depth information cannot obtain dense and accurate depth information at low cost.Therefore,completing a single color picture or a sparse depth map into a dense depth map has always been a popular research topic of research significance in the field of computer vision and pattern recognition.In this paper,deep learning methods will be used to study the monocular depth estimation and depth completion tasks,and the main works are:(1)In this paper,a monocular depth estimation algorithm based on multi-feature fusion is proposed.The algorithm adopts an end-to-end encoding and decoding structure,in the encoding branch,by increasing the residual channel attention module,features of different sizes can be generated from shallow to deep,which can effectively improve the accuracy and reliability of the encoding,and can adaptively learn the importance of each channel as needed,and retain the important feature channels to remove redundant information,thereby improving the expression effect of the feature map.At the same time,the module uses dilated convolution to further expand the receptive field,extract local spatial features,and extract global feature information while ensuring that the gradient will not explode.The coding features and pooling operations of different scales are introduced into the decoding branch to fuse the spatial dimension and channel dimension of the decoding features of different scales,and integrate them into a unified feature map to achieve a more accurate feature representation and eliminate the semantic gap between features,and further extract implicit spatial geometric features.(2)This paper proposes a simple and effective deep completion algorithm that aims to combine the advantages of traditional image processing techniques and popular deep learning methods.This method uses a two-branch backbone network to improve the existing backbone network and enhance the effect of color information and deep information feature fusion.At the same time,the algorithm contains two modules using traditional image processing technology: adaptive densification module and coordinate projection module.The Adaptive Densification module adaptively completes low-gradient areas in a sparse depth map to generate semi-dense depth maps.Semi-dense depth maps contain denser geometric information and enhance the robustness of shallow features in convolutional neural networks.The Coordinate Projection module takes a semi-dense depth map as input and projects it onto a 3D pose map.Enhanced 3D topology and geometric relationships in depth maps. |
