Fast Semantic Segmentation of 3D Lidar Point Cloud Based on Random Forest Method

Fast semantic segmentation of the 3D lidar point cloud is useful for 3D reconstruction, mapping, and unmanned automobile. The challenges of semantic segmentation for the existing deep learning method require large amount data and high computational cost and hardware. This paper adopts a machine learning method based on random forest with few samples, which can process highdimensional features in parallel for 3D lidar point cloud to achieve fast semantic segmentation. First of all, the proposed framework will analyze position, intensity and color to extract twenty features such as eigenentropy, verticality, average intensity and height statistics. Secondly, the feature representation of multi-scale neighborhood is constructed by grid downsampling and K-nearest neighbor (KNN) algorithm. Thirdly, the optimized feature set is re-input to the random forest classifier to achieve semantic segmentation based on the random forest feature selection (RFFS) process.The proposed method has been tested on the high-precision 3D lidar point cloud dataset of Fudan University Jiangwan Campus and the public dataset Toronto-3D. The results show that semantic segmentation can be achieved well. For roads, lane marking, vegetation, buildings, pedestrians, vehicles, and poles in the high-precision 3D lidar point cloud dataset of Fudan University Jiangwan Campus, the mean intersection over union (MIoU) and the overall accuracy (OA) are greater than 90%, the test results in the public dataset set Toronto-3D are better than pointnet++. The method described in this paper has a fast computation speed than the deep learning method, which may be applied to handle millions of lidar 3D point clouds more quickly in the future.