Monocular Image Depth Estimation Based On Unsupervised Learning

The acquisition technology of depth information is widely used in the fields of 3D reconstruction,scene perception,unmanned driving,visual navigation and obstacle detection.In some professional scenarios,it is very difficult to obtain high-density and accurate depth information.Relatively speaking,the method based on deep learning does not require expensive equipment,and the operation is relatively simple,and the scope of application is wider.At present,the mainstream monocular image depth estimation is mostly based on supervised learning,which requires a large-scale ground truth data set,and the real depth data per pixel is difficult to obtain.In addition,although the monocular depth estimation based on unsupervised learning can reduce The requirements for the data set,but will cause problems such as unclear boundary contours of object depth information and object occlusion artifacts.In response to the above problems,an unsupervised monocular image depth estimation method with a joint attention mechanism is proposed,which can obtain more accurate depth estimation results.The main contents include:(1)Aiming at the problem that large-scale ground truth data sets are difficult to obtain,based on an unsupervised method,it is proposed to use the Attention-Unet network architecture with jump connections in the depth estimation network structure to evaluate the importance of features at different spatial locations.Control,enhance the resolution of object features,and improve the accuracy of depth estimation.(2)Aiming at the problems of unclear object boundary contour,object occlusion,artifacts,etc.in the method based on unsupervised monocular depth estimation,it is proposed to introduce a self-attention module in the Attention-Unet network codec to re-encode the encoder The output features are processed,and the self-attention mechanism is realized by supervising the features at the upper and lower levels,thereby improving the definition of the object boundary contour.In addition,the use of minimized luminosity reprojection loss and automatic masking loss,using the minimum value of the source image error on each pixel for processing,solves the problem of object occlusion and reduces object artifacts.(3)Aiming at the current research on the depth estimation of monocular images based on deep learning,it is difficult to directly obtain specific distance values from experimental results,and proposes a vehicle measurement that combines the unsupervised learning monocular image depth estimation model with the YOLOv3 model.The distance method realizes a vehicle-oriented distance measurement simulation system.The algorithm in this paper trains the network model on the KITTI dataset and tests it on the Eigen Split dataset.Experimental results show that compared with the current mainstream methods,the algorithm proposed in this paper can improve the accuracy of distant objects while making the contours of objects in the renderings clearer,and the results are more accurate.In addition,the vehicle ranging method based on monocular vision proposed in this article can be effectively applied to vehicle detection,ranging and other auxiliary driving systems.