City-Scale Street Trees Mapping through LiDAR, RGB Camera, and INS Integrated Vehicle Sensor System.

Vehicle-installed light detection and ranging (LiDAR), RGB camera, and inertial navigation system (INS) enable effective city-scale mapping of street trees which includes structural parameters, species, and location. To map street trees, extracting the trees from the point cloud is a core process in that range information of the target is essential for the mapping. Although recent research on point cloud deep learning showed that object detection within the point cloud is feasible, applying it to the street tree is challenging because constructing the training data costs a lot. In this study, we propose a novel method that extracts individual tree point clouds by integrating point clouds and RGB imagery, which is cheaper training data compared to the point cloud. To this end, binary image segmentation of street trees by U-Net is done and the ground-removed point cloud is projected to the segmented image so that the street tree point cloud candidates are selected. It is noteworthy that the species classification results by Yolo_v3 are linked with the street tree candidates from this stage. Since the candidates often involve non-tree objects or multiple objects in a single candidate, a principle component analysis (PCA) and Voronoi diagram are used to filter them out from the candidates. Once the individual street tree point clouds are prepared, the structural parameters are calculated. Finally, a street tree map is created by transforming tree point clouds to the geographic coordinate from the sensor's local coordinate. The proposed method showed a robust performance in various street settings all over Suwon city, Korea. This study implies that the vehicle sensor system can automate the mapping process of street trees, which was a time-consuming and costly task with conventional field surveys.