Improving Accuracy of Pseudo-LiDAR for 3D Object Detection by Accurate Depth Estimation

With the advancement of autonomous driving technology, traditional 2D object detection approaches no longer fulfill the Safety of The Intended Functionality (SOTIF) standards. LiDAR sensors are efficient in giving precise point cloud information for accurate and reliable 3D object detection. Alternative sensors also are researched by researchers if they have lower costs but are limited in environmental circumstances. Notably, pseudo-LiDAR-based monocular 3D object detection has emerged as a viable option, as it generates a depth map from an RGB picture before translating it into a point cloud for 3D object detection. However, because the majority of depth estimation algorithms do not account for the loss of boundary information during feature extraction processing, depth artifacts occur in the periphery of objects in the depth map. These artifacts introduce long-tail problems in pseudo-LiDAR data, undermining the accuracy of 3D object detection. Therefore, we suggested a depth estimation method using Laplacian pyramid-based depth residuals to correctly capture object depth bounds. This improved the estimate and revised the depth map, which was converted into pseudo-LiDAR point cloud data. The pseudo-LiDAR was then used to recognize 3D objects. The proposed method, in particular, significantly mitigated the common border Long-tail problems in pseudo-LiDAR data, hence improving the precision of 3D object detection. Experiment validations demonstrated the method's efficiency and enhanced performance to improve the reliability of autonomous driving systems.