ROAD TYPE CLASSIFICATION OF MLS POINT CLOUDS USING DEEP LEARNING

Abstract. Functional classification of the road is important to construction sustainable transport systems and proper design facilities. Mobile laser scanning (MLS) point clouds provide accurate dense 3D measurements scenes, while their massive data volume lack structure also bring difficulties in processing. cloud understanding through deep neural networks achieves breakthroughs since PointNet arouses wide attention recent years. In this paper, we study automatic type MLS by employing a point-wise network, RandLA-Net, which designed for consuming large-scale clouds. An effective local feature aggregation (LFA) module RandLA-Net preserves geometry formulating an enhanced geometric vector learning different weights neighborhood. Based on method, investigate possible combinations calculate neighboring weights. We train colorized from city Hannover, Germany, classify points into 7 classes that reveal detailed functions, i.e., sidewalk, cycling path, rail track, parking area, motorway, green island without traffic. Also, three inside LFA are examined, including only, combined with additional features (e.g., color), differences features. achieve best overall accuracy (86.23%) mean IoU (69.41%) adopting second third respectively, Red, Green, Blue, intensity. The evaluation results demonstrate effectiveness our but observe types benefit most settings.