Research On Monocular Depth Estimation Based On Self-supervised Learning

Monocular depth estimation aims to infer per-pixel depth from a single RGB image,which has been widely applied in autonomous driving,augmented reality and3 D reconstruction.In recent years,deep learning-based methods for monocular depth estimation have made promising advancement.Supervised learning-based monocular depth estimation methods generally rely on large datasets annotated with ground-truth depth to train depth estimation model.However,in practice,obtaining pixel-wise depth annotation with high quality for a large number of images is a challenging task,which greatly limits the application of supervised methods.Self-supervised learning-based monocular depth estimation methods utilize different view images to provide supervision signals for training a network without relying on real depth label.However,how to effectively use different view images to provide powerful supervision and constraints for the monocular depth estimation model is still a challenging task.Based on the in-depth analysis of the key issues of monocular depth estimation,with the aim of improving the accuracy of monocular depth estimation,this thesis has carried out research on self-supervised learning-based monocular depth estimation technology.This thesis proposes a self-supervised monocular depth estimation method based on predicted binocular cue and occlusion-guided constraint.First,a binocular cue prediction module is presented to generate auxiliary vision cue for providing further supervision and constraint to monocular depth estimation network by effectively exploring the geometric correlation between the source view and the synthesized target view.In addition,in order to address occlusion problem,the occlusion region is inferred by utilizing the generated vision cue,and an occlusion-guided constraint is developed to balance the reconstruction error between the occlusion region and the non-occlusion region,so as to better guide the optimization of the network.Experimental results on two popular benchmark datasets have demonstrated that the proposed method significantly improves the quality of self-supervised monocular depth estimation and achieves comparable results to supervised monocular depth estimation methods.This thesis also proposes a self-supervised monocular depth estimation method based on monocular completion assistance and knowledge transfer.First,in order to effectively use the correlation between the two tasks of monocular depth completion and monocular depth estimation,the monocular depth completion task is proposed as an auxiliary task to further enhance the performance of self-supervised monocular depth estimation.Moreover,in order to maintain a self-supervised learning paradigm,the traditional stereo matching algorithm and a random point sampling strategy is used to construct training data for the monocular depth completion task.In order to make the monocular depth estimation network benefit from the effective feature representation learned by the monocular depth completion network,a knowledge transfer strategy is proposed,which transfers the related knowledge learned by the monocular depth completion task to the monocular depth estimation network.Experimental results on widely used benchmark datasets show that the proposed method achieves better depth estimation results.