SASDC: Semantic-Assisted Self-Supervised Depth Completion with Stereo Camera and Single-Line LiDAR

In autonomous driving and robot navigation, stereo camera-based depth estimation methods are prone to errors in low-texture or long-distance scenes, making it challenging to meet the needs of complex environments. Although multi-modal methods incorporating multi-line LiDAR can effectively address the limitations of vision-based methods, the high cost and computational complexity of multi-line LiDAR pose significant challenges to practical applications. To solve the issue, we propose a self-supervised depth completion method that achieves depth completion for single-line LiDAR and stereo camera through feature fusion, pre-trained monocular depth estimation model for supervision, and semantic-assisted perception. First, independent encoders are utilized to separately extract features from both RGB images and single-line LiDAR depth maps, thereby achieving depth completion through feature fusion. Then, a pre-trained semantic encoder guides the RGB encoders to capture higher-dimensional continuous semantic information. Finally, the relative depth map generated by a pre-trained monocular depth estimation model is employed for self-supervised training. Extensive experiments demonstrate that the proposed method can: accurately reconstruct complete depth maps from single-line LiDAR depth maps and RGB images, and effectively reduce depth errors in regions with weak textures.