| Depth estimation is a key technique for 3D reconstruction,environment perception and viewpoint synthesis.Depth information is of great importance for computers to understand the geometric relationships of a scene.Computer understanding of scene geometry depends on depth information.Laser radar ranging is a relatively accurate method to obtain depth information,but the emission equipment of laser radar is too expensive.There are two kinds of methods for the depth estimation task based on image processing,namely binocular depth estimation and monocular depth estimation.The binocular depth estimation has obvious disadvantages,such as large size of hardware and expensive computation,while the monocular depth estimation does not.However,the images taken from different range of visibility may be same,the monocular methods have the problem of the ambiguity of the depth value,which will lead to some concerns,such as 1)obvious object loss;2)blurred edge details of adjacent depth.In recent years,the development of depth learning has greatly improved the performance of depth estimation.However,the monocular depth estimation model has problems such as poor performance in complex environments,and the model is too sensitive to light.Based on the monocular depth estimation method,this paper has carried out the following research work in view of the current problems in depth estimation:(1)A monocular depth estimation algorithm based on the fusion of scene object attention mechanism and weighted depth map is proposed.In this part,a scene object attention module is designed to enhance the ability of the model to obtain context information by calculating similar feature vectors between different locations,thus reducing or eliminating the problem of incomplete object depth information in natural scenes.At the same time,we design a weighted depth map fusion module based on multi-vision granularity,in which the 1 × 1 convolution layer is adopted to compute the multi-scale coarse depth map’s weight learned from standard back-propagation.Multivision granularity depth maps with different depth information are assigned different weights for fusion.The fused depth map contains depth information and rich scene object information to effectively solve the problem of detail blur.The proposed network is trained,verified and tested on KITTI.The results show that the proposed a monocular depth estimation algorithm based on the fusion of scene object attention mechanism and weighted depth map achieves an accuracy of 0.879 and a square relative error of 0.765 and a root mean square logarithmic error of 0.185 in the case of σ<1.25.The prediction performance of this algorithm on DDAD data set is also significantly better than other depth estimation algorithms.(2)A depth estimation algorithm based on illumination insensitivity and semantic information fusion is developed.On the basis of the previous work,in this part,we design a low-light image enhancement module to reduce the impact of light or complex environment on depth estimation,by improving the image quality of the training dataset to reduce the sensitivity of the model to light.At the same time,pixel adaptive convolution is used to obtain semantic depth features from the pre-trained semantic network,and the prior knowledge and auxiliary information of the scene are used to guide the depth estimation model for depth estimation,so as to improve the prediction performance of the model.In addition,the atrous spatial pyramid pooling module is introduced to refine the semantic scene level,increase the receptive field of the network,and improve the ability of the network to obtain feature information,which could solve the problem of depth map edge ambiguity.The experimental results show that the algorithm can improve the generalization of the depth estimation model and achieve the overall improvement of the depth map accuracy.The root mean square error(RMSE)reaches 4.150.In summary,aiming at the problems of blurred details of adjacent depth edges and low prediction accuracy of monocular depth estimation algorithm,a weighted depth map fusion module and a method of integrating semantic information from the perspective of improving the ability of network to obtain information are proposed.In order to improve the obvious problem of missing objects,a scene object attention mechanism is developed.In order to reduce the impact of light or complex environment on the performance of depth estimation,we use the low-light image processing model to enhance the trained dataset from the perspective of improving the image quality of the low-light images in the dataset. |
