TriSpaSurf: A Triple-View Outline Guided 3D Surface Reconstruction of Vehicles from Sparse Point Cloud

Vehicle is one of the important subjects studied in the domain of computer vision, autonomous driving and intelligent transportation system. 3D models of vehicles are used widely in the literature of vehicle categorization, pose estimation, detection, and tracking. However, previous work uses only a small set of 3D vehicle models either from CAD design or multi-view reconstruction, limiting their representation ability and performance. A feasible approach to acquire extensive 3D vehicle models is desired. In this paper, we are interested in 3D surface reconstruction of on-road vehicles from sparse point cloud captured by laser scanners equipped ubiquitously on autonomous driving platforms. We propose an innovative reconstruction pipeline and method, called TriSpaSurf, which could reconstruct unbroken and smooth surface robustly from just a single frame of noisy sparse point cloud. In the TriSpaSurf, triple-view 2D outlines are first fitted on the 2D points from the projection of 3D point cloud under each view, and then 2.5D surface reconstruction is carried out under the guidance from triple-view outlines. By projecting 3D point cloud onto 2D views, 2D outlines could be estimated robustly due to the reduced complexity and higher signal-to-noise ratio in 2D views, and could provide fairly stable and tight multi-view constraints for 3D surface reconstruction. The effectiveness of our method is verified on the KITTI and Sydney dataset.