Research On Unsupervised Monocular Depth Estimation Method For 3D Object Detection

With the development of technology and progress of industry,until now,AI has become a hot topic and a hot research direction in the information age.Its main core technology,deep learning,and its landing application,automatic driving system,are the frontier fields that are currently attracting our focus.This paper mainly discusses the vision problem of the perception module in the autonomous driving system.The top priority of autonomous driving is safe driving,which requires the vehicle be able to “see” objects and measure the distance to them.Therefore,reliable and stable object detection,fast and accurate depth estimation have become the basic requirements for autonomous driving.Meanwhile,depth estimation can be regarded as an upstream task of object detection,and the PseudoLi DAR method fills the gap between depth estimation and 3D object detection.In reality,in the face of expensive Li DAR sensors and huge costly human-labeled datasets,pure image-based methods still deserves expectation,and the research on unsupervised learning also has a long way to go.In view of the above practical needs,this paper proposes an unsupervised monocular depth estimation method based on pure images for 3D object detection.The main work is provided as follows:1.The article proposes a general point cloud consistency constraint method,and correspondingly design the loss functions for monocular depth estimation.Previous monocular depth estimation methods often pay more attention to the RGB information and ignore 3D structural information.To this end,based on the principle of multi-view geometry and that 3D scene reconstruction,and also considering combining the pose changes of cameras between temporal image sequence to link the 3D reconstructed point clouds of different frames,this paper proposes the idea of point cloud consistency constraint.With the idea,this paper designs corresponding loss functions from the perspective of geometric absolute distance,geometric relative distance,and structural similarity of point clouds,and finally proposes a integrated loss function based on point cloud consistency constraints.Through experimental verification,compared with the benchmark model Mono Depth2,the introduction of point cloud consistency constraints helps improve the accuracy of depth estimation,which facilitates downstream 3D target detection.This paper also verifies the generality of point cloud consistency constraints.For several unsupervised monocular depth estimation models that take temporal image sequences as input,the point cloud consistency constraint is demonstrated applicable too.2.The article Proposes an unsupervised monocular depth estimation method based on point cloud learning.The previous work relies on the Pseudo-Li DAR generation algorithm,that is,the depth map predicted by the depth estimation model is firstly converted into Pseudo-Li DAR,which is then used to train the 3D target detection model.To solve this problem,a point cloud estimation model Mono PC based on the point cloud decoder PC-Decoder is proposed.According to the characteristics of point cloud learning tasks,this paper designs a point cloud consistency constraint scheme for Mono PC,and correspondingly proposes a self-supervised constraint loss for predicting point clouds and an interframe constraint loss for predicting point clouds.Compared with the benchmark model Mono Depth2,the experimental results of depth estimation and 3D object detection both prove the effectiveness of this work.