Robust Ground Segmentation in 3D Point Cloud for Autonomous Driving Vehicles

3D LiDAR sensor is a key sensor for autonomous driving, and ground segmentation based on point cloud is a fundamental and essential task for various purposes. Although ground segmentation achieves relatively good result based on plane fit, but misclassification often occurred in the intersection between ground and vertical objects or place due to inappropriate ground plane fitting threshold. Therefore, this paper proposed an improved algorithm based on Patchwork++ to obtain more robust ground segment result. In this paper, we first estimate the segmented section planes based on Patchwork++, and then further evaluate the estimated ground points using our proposed intersection re-evaluation method, then the points were mapped to an image plane with three channels, and assigned to different channel according to their classification, ground points assigned to green channel and non-ground points assigned to red channel, the G and R channels are then dilated and the overlapping regions are reclassified by jumping convolution operations, finally the classification results are reflected back to the 3D point cloud. By testing on the SemanticKITTI dataset, the experimental results showed that our method achieves 96.5% F1-score when the vegetation is excluded, which is 0.6% higher than original method, and achieves 94.2% F1-score when the vegetation is added, is 1.1% higher than original method.