Road classification using 3D LiDAR sensor on vehicle

Abstract

Road classification is essential in outdoor vehicle perception systems to evaluate driving scenarios. A light detection and ranging (LiDAR) sensor can provide accurate measurements of an environment composed of range, intensity, and angular information. This study proposes a weakly supervised learning algorithm for classifying road surface types based on laser measurements. The proposed algorithm first separates the laser points on the road surface based on the angular structure of the initial range measures obtained using 3D LiDAR in a noniterative manner. Then, a physical-data-driven model that represents the retroreflective nature of the different surface types is abstracted from the intensity measures and their factors. The linear regression parameters of the model are employed as feature descriptions of the road surface. The K-means method is used to classify the different road types. The learning process requires weak supervision at the training centroid stage and is robust to new road types. The classification performance was evaluated in different outdoor scenarios, and the classification accuracies exceeded 98%. The proposed algorithm outperforms similar existing methods, performs at the same frequency as that of LiDAR, and can be implemented in real time.