A Knowledge-Based Multi-Scale Adaptive Classification Approach for Mobile Laser Scanning Point Clouds in Urban Scenes

With the quick development of mobile light detection and ranging (LiDAR) systems, point clouds are frequently applied for various large-scale outdoor scenes. It is fundamental to quickly and accurately classify objects of mobile laser scanning (MLS) point clouds in such urban scene applications. However, an important problem is the need for massive training samples in object classification. High computational cost is also a common challenge. To overcome them, a knowledge-based multi-scale adaptive classification approach (KMAC) is proposed in the paper. The method consisting of four layers derives from a normal neural network framework, the operation in part layers differ. As the scale difference of various objects in natural environment, 3D multi-scale spatial local relation of objects is explored with inspiration by the idea of convolution. Two types of distinguishable features of actual objects are explored to describe 3D point clouds by a 2D vector representation. Then, human knowledge is used to directly build an end-to-end match between these feature descriptions in 2D and 3D point clouds of actual objects. Point clouds which are adjacent with the same feature representation would be intentionally integrated into multiple adaptive regions. The adaptive integration solves scale difference of various objects. The direct match by knowledge exactly plays the role of training samples. Qualitative and quantitative experimental results on three data-sets finally show the proposed approach is promising to efficiently classify unlabeled objects in urban scenes.

Automated summary

In this paper, a knowledge-based multi-scale adaptive classification approach is proposed for efficiently classifying objects in mobile laser scanning point clouds. By exploring the multi-scale spatial local relation and distinguishable features of objects, the method establishes a match between the feature descriptions and actual objects' point clouds based on human knowledge. Experimental results demonstrate the promising performance of the proposed approach for classifying unlabeled objects in urban scenes.