Classification of LiDAR point clouds using supervoxel-based detrended feature and perception-weighted graphical model

Abstract

Interpretation of 3-D scene through LiDAR point clouds has been a hot research topic for decades. To utilize measured points in the scene, assigning unique tags to the points of the scene with labels linking to individual objects plays a crucial role in the analysis process. In this article, we present a supervised classification approach for the semantic labeling of laser scanning points. A novel method for extracting geometric features is proposed, removing redundant and insignificant information in the local neighborhood of the supervoxels. The proposed feature extraction method uses the supervoxel-based local neighborhood instead of points as basic elements, encapsulating the geometric features of local points. Based on the initial classification results, the graph-based optimization is used to spatially smooth the labeling results, based on the graphical model using the perception weighted edges. Benefiting from the graph-based optimization process, our supervised classification method required only a few training datasets. Experiments were carried out by comparing the semantic labeling results with manually generated ground truth datasets. The performance of the proposed methods with different characteristics was analyzed. By using our testing datasets, we have achieved an overall accuracy of better than 0.8 for assigning the measured points to eight semantic classes.