BUILDING DETECTION FROM AERIAL LIDAR POINT CLOUD USING DEEP LEARNING

With the development and widespread application of aerial LiDAR, point cloud data can easily be acquired and used in many fields. The accurate detection of buildings from an aerial LiDAR point cloud has attracted much attention owing to its wide range of applications, such as updating building maps and constructing 3D city models. However, such applications remain challenging in the fields of photogrammetry, remote sensing, and computer vision. In this paper, we discuss the features that contribute to building detection accuracy from an aerial LiDAR point cloud using a deep-learning-based method (KPConv). We evaluated the influence of neighborhood size, intensity, RGB, and normal vectors on building detection. The study area was approximately 6 km2, consisting of 133 million points and 8,099 buildings. The density of the point cloud data was eight points/m2. We compared search radii of 4, 10, 25, and 50 m for finding neighboring points. The results suggest that an optimal neighborhood size improves the accuracy of building detection. For searching neighboring points, a radius of 25 m is optimal when the building area is less than 1000 m2, whereas a radius of 50 m is optimal when the building area is larger than 1000 m2.We also compared different features as inputs to KPConv for training and testing, such as i) 3D coordinates only, ii) 3D coordinates and intensity, iii) 3D coordinates and RGB, iv) 3D coordinates and normal vectors, and v) 3D coordinates, intensity, RGB, and normal vectors. The results suggest that neither intensity nor normal vectors contribute to the accuracy of building detection, while the features of RGB have a limited effect on the results.