Enhancing Depth Image with Richer Convolution Feature for Improved Perception

Depth cameras are widely used in various fields, particularly in robotics, for tasks such as robot path planning, object avoidance, target tracking, and autonomous driving. However, the incomplete depth sensing technology of depth cameras results in depth images with significant noise, including errors in depth values and detection failures due to environmental instability. This noise adversely affects the performance and stability of robotic applications. To address this issue, various filtering methods have been proposed, but the corrected depth images still contain residual noise. This paper utilize a deep learning-based edge detection method, Richer Convolution Feature, to accurately estimate planar regions in depth images. The RANSAC algorithm is used to enhance the planes of objects. Experimental results demonstrate the effectiveness of the proposed approach in reducing noise and improving depth image quality.