Automatic classification of mobile laser scanning point clouds in urban road environments is a fundamental yet challenging problem for the full exploitation of related data sets. Deep learning networks originated from computer vision have recently demonstrated a high potential for three-dimensional data classification in these complex scenarios. However, directly processing massive three-dimensional points with deep learning networks often fact the problems of unorganised structure and differences in scenarios. Thus, this paper presents a method based on a three-dimensional deep learning network that directly classifies raw point clouds in urban road environments. This method consists of a symmetric ensemble point (SEP) network designed by applying a symmetric function to capture different scales of those relevant features, and by selecting the optimal sub-samples using an ensemble method. The experimental results indicate that this method effectively distinguishes six types of objects: roads, buildings, walls, traffic signs, trees, and light poles. The achieved average classification accuracy is approximately 96.93%, which is suitable for practical use in transportation network management.

CICTP