Self-Supervised Monocular Depth Estimation With Depth–Motion Prior for Pseudo-LiDAR

In recent years, learning-based depth estimation has advanced significantly due to multiview geometry and self-supervised methods. These approaches use video sequences or multiview images for supervision, making training data easier to obtain. However, challenges remain, including the need for extrinsic camera precalibration and insufficient real-time performance. To address these issues, we propose depth-motion prior (DMPDepth), a self-supervised monocular depth estimation method with multiview stereo (MVS) prior. This approach integrates the computational efficiency of monocular depth estimation with the accuracy of MVS matching. We incorporate a pose estimation module that eliminates the need for camera extrinsic precalibration. In addition, a depth motion prior is introduced by combining triangulation with camera motion, and a depth motion prior guidance strategy is proposed to improve the real-time performance and accuracy. We demonstrate the significantly superior performance of DMPDepth over other public methods on the KITTI and self-constructed Campus Road datasets, with reduced inference time. Furthermore, a pseudo-LiDAR system for autonomous driving based on DMPDepth successfully demonstrates accurate real-time depth estimation in a real Campus Road environment.