Local Linear Fitting Method for Lidar Depth Completion Based on RGB Image Guidance

The depth image contains rich spatial information, which is crucial for simultaneous localization and mapping (SLAM) as well as accurate environmental perception. However, the original depth image is often sparse and contains small holes with missing depth values. In this paper, we have implemented point cloud depth image projection and proposed an RGB image guided completion method. Objects of the same class typically have similar color and depth values. Our local linear fitting method utilizes regions with similar colors to construct a completion objective function. Experiments in both indoor and outdoor scenes demonstrate that our method can generate a depth image with the same resolution as the RGB image. Analysis reveals that our method enhances the quality of depth image completion compared to traditional methods that rely solely on point cloud kernel dilation.