Research On Monocular Depth Estimation Based On Deep Learning

With the highly developed of technology and science,intelligent robots have entered our lives and brought a lot of convenience to our daily life.Intelligent robots can replace humans to do some extreme or dangerous works,so it need autonomous navigation and obstacle avoidance technologies.SLAM(Simultaneous Localization and Mapping)algorithm can help realize intelligent robot navigation,positioning,mapping,and other functions,is an integral part of the intelligent robot.In recent years,visual SLAM algorithm has become increasingly mature,and it is difficult to make big theoretical breakthroughs.Deep learning technology is changing rapidly,and great progress has been made in many aspects.Therefore,people focus on the combination of deep learning technology and visual SLAM algorithm,and get good results.In terms of camera,monocular camera has low cost,simple structure and is suitable for market application in a variety of scenes.Therefore,this paper uses monocular camera and improves the ORB?SLAM2algorithm,which is one of the widely used algorithm in the industry.In view of the above situation,a deep learning based monocular SLAM algorithm is implemented in this paper.The research work is summarized as follows:(1)In view of the situation that the camera image is blurred and blocked by the rain line in rainy days,this paper proposes an algorithm of single image deraining and depth estimation based on deep learning,and realizes the deraining and depth estimation of images simultaneously by using the multi-task learning method.An algorithm of encoder-decoder structure is implemented based on convolutional neural network in this paper,the decoder part use deconvolution methods respectively regression atmospheric optical map,direct deraining map,rain streak map,transmission map and direct depth estimation map.After that we can get the final deraining map and direct depth estimation map.The final depth estimation map is obtained by cascading information fusion with the direct depth estimation map and the final deraining map after edge extraction by Canny operator.In the aspect of loss function,depth and transmission consistency loss are added to improve the effect of image deraining and depth estimation.We synthesize NYU?rain and Outdoor?rain datasets and experiments are carried out on them.The accuracy of depth estimation on images with rain is better than that of the existing algorithms.In terms of running time,the algorithm of depth estimation and image deraining simultaneously in this paper is faster than that of the existing algorithms of image deraining and depth estimation respectively.(2)In view of the difficulties faced by the monocular SLAM algorithm,a monocular SLAM algorithm based deep learning is designed and implemented in this paper,and based on the dense depth map obtained by deep learning,the absolute scale depth information is obtained,which solves the problems of scale uncertainty and drift.The overall framework of the algorithm and the concrete implementation steps of initialization,motion estimation and dense point cloud reconstruction are introduced.Combined with the dense depth map obtained by deep learning,the absolute scale is confirmed during initialization.In order to reduce the amount of calculation and speed up the algorithm,only to estimate the depth information of key-frame.RANSAC algorithm is used to correct the scale of the non-key frame through the depth information of key-frame during the motion estimation.We assuming that the inverse depth conforms to the Gaussian distribution,and use the method of polar line search and block matching to reconstruct the non-key frame dense depth map.Then reconstruct the dense point cloud map by the dense depth map of all frames.Therefore,the trajectory information and scene information are obtained at the same time.Experiments on the TUM dataset have been conducted and good results have been obtained.