Research On Unsupervised Monocular Depth Estimation Algorithm Based On Video Sequence

Monocular depth estimation refers to obtaining distance information from each point in an image to the camera based on a single RGB image,which is also known as depth information.In many applications such as autonomous driving,3D reconstruction,and object tracking,depth information plays an extremely important role.At present,there are two methods for monocular depth estimation: geometric based methods and deep learning based methods.The monocular depth estimation method based on depth learning uses end-to-end training,which is more efficient than the traditional geometry based method,and can adapt to different scenarios and task requirements through transfer learning or Domain Adaptation and other methods on different data sets,with a wider applicability.Therefore,the monocular depth estimation method based on deep learning has become the current mainstream research direction.Monocular depth estimation methods based on deep learning can be divided into supervised methods and unsupervised methods.The supervision method requires a large number of datasets with depth labels as training data,which has high prediction accuracy.However,the cost of labeling depth label data is high,making it difficult to obtain a large amount of training data.Unsupervised methods do not require depth labels and train the model by minimizing reconstruction errors,smoothness,geometric constraints,and other factors related to image and depth,reducing dependence on depth labels.However,the prediction accuracy may be lower than that of supervised methods.This article proposes a monocular depth estimation algorithm DsMDE(Domain separation Monocular depth estimation)based on domain separation feature adaptation to address the limitations of supervised and unsupervised methods.The main contributions of this method are as follows:(1)Improving the Res Net based depth estimation framework,using the HR-Net+GAU(Global Attention Upsampling Module)structure as the depth estimation framework.This network structure has more refined feature extraction and processing capabilities,improving the prediction accuracy of unsupervised algorithms.(2)The supervised training of synthetic data is combined with the unsupervised training of real data,and the domain separation and maximum mean difference in Domain Adaptation are introduced to eliminate the domain difference between synthetic data and real data,thus improving the unsupervised algorithm accuracy.The experimental results on the KITTI dataset show that the HR-Net+GAU depth estimation network structure used in the DsMDE method in this paper can reduce the absolute relative error of the unsupervised algorithm to 0.104 and improve it by up to 10%;After combining supervised training with unsupervised training,and eliminating the difference between synthetic data and real data through Domain Adaptation,the absolute relative error decreases to 0.100,the square relative error decreases to 0.672,and other indicators are improved compared with unsupervised methods.This result proves the effectiveness of the DsMDE algorithm in this paper.