A Consideration on Automatic Identification of Road Features for Area Data Generation of Point Cloud Product Model

Abstract

With the innovation of laser measurement equipment, point cloud data are being measured. However, since each point in a point cloud has no meaning, one way to handle point cloud data wisely is through product modeling, in which the meaning of the road feature is assigned to the point cloud data and structured. To solve this problem, area data generated by reference to the outline and position coordinates of the road feature is needed. The authors proposed an area data generation method using HD maps for automated driving, but area data cannot be generated without HD maps. In this study, we compared the identification accuracy by utilizing ConvPoint and RandLA-Net, techniques for automatically identifying road features, to generate area data by identifying objects even outside the areas covered by HD maps. As a result, the average F-measure of ConvPoint was 0.927. Contrarily, RandLA-Net built a model three times faster than ConvPoint, with an average F-measure of 0.852.