Deep Learning Based Monocular Video Depth Estimation And Ego-motion Prediction

Nowadays,with the development of 3D techniques,there is an increasing requirement for depth.For examples,depth plays an important role in automotive industry and the processing of 3D reconstruction.And in 3D volumetric display,depth contributes to the stereo vision.However,the depth estimation is an ill-posed problem when the input is monocular images or videos from which the monocular depth estimation is based on the geometry structure.Although neural networks improve the accuracy of depth estimation,the collection of ground truth,which is necessary in the supervised training,is costly.As a result,the unsupervised training is employed in the depth estimation.The accuracy of view synthesis can be used to train the network in unsupervised method.The unsupervised depth estimation of monocular video takes monocular image or video sequence as input and passes through the depth estimation system.Then,the depth can be predicted from the boundary,fog effect and positional relationship hidden in the optical images.The influence of the smooth constraint used in the estimation process on the depth estimation will cause the boundary of the depth image to be blurred.Aiming at the problem of blurring,the detailed processing of depth pictures and the self-supervised process of network training are studied,and an improved method is proposed.The specific work is as follows.In order to improve the accuracy of smoothness,the dynamic filter network is employed to adaptively filter the depth image and bridges the depth image and optical image.In the experiment,comparing to the structure from motion the accuracy of depth estimation is improved by 1.35%,there is no obvious influence on pose estimation.The super-resolution accuracy is used as the self-supervision in the depth estimation network,where the comparison of super-resolution image and large image is concatenated to the feature map.The combination of super-resolution results is weighted by the variation of depth image,as a result,the accuracy of super-resolution is related to the depth image.The result of depth estimation is improved by 3.36%.