Urban Road Boundary Modeling Via Multi-source Data

Mobile laser scanning(MLS),used for 3D city modeling,has become a cost-effective solution to capture various dense point clouds with high precision.Considering that road boundary requires relatively high accuracy and completeness and the point clouds acquired by MLS have the advantage of high-density,long range,and high-efficiency,the system is suitable to be used to extract road boundary with high accuracy.However,due to occlusion during collection and data density decreasing with distance increasing,the raw point clouds are always incomplete,resulting in the incomplete extraction result of road boundary.For the imcomlete problem of extracted road boundary,this paper presents a deep learning-based framework for constructing complete 3D urban road boundary using multi-source data,which include mobile laser scanning(MLS)point clouds,spatial traj ectory data,and remote sensing images.The proposed method first extracts 3D road boundaries based on MLS point clouds.Then,road boundaries are completed by image-based gap detection and inpainting using convolut:ion neural networks.To solve the uncertainties of gaps,road centerlines generated from dynamic taxi satellite positioning trajectory data and remote sensing images are used as completion guidance for a conditional generative adversarial nets model to obtain more accurate and complete road boundaries.Finally,after associating a sequence of satellite positioning recorded trajectory points with the correct 3D road boundaries,inherent geometric road characteristics and road dynamic information are extracted from the complete boundaries and taxi satellite positioning traj ectory data,respectively.The proposed method is thoroughly tested on self-test data(MLS point clouds)and the internationally recognized KITTI data.The completeness,correctness,and quality of the proposed completion method achieved using the three datasets,are 93.78%?95.38%?89.71%,respectively(the international conference and exhibition center dataset),93.13%?96.22%?89.84%,respectively(the coastal ring road dataset),and 94.41%?86.82%?82.57%,respectively(KITTI dataset).In addition,the boundary refinement modules for gaps without completion and irregular completion structures achieved results with an average error distance of 0.48 m and 0.56 m,respectively.The experimental results on point clouds from different sensors demonstrate the proposed method achieves excellent results for constructing complete 3D urban road boundary and extracting road characteristics.Its application is very wide.