Diffusion and inpainting of reflectance and height LiDAR orthoimages

This paper presents a fully automatic framework for the generation of so-called LiDAR orthoimages (i.e. 2D raster maps of the reflectance and height LiDAR samples) from ground-level LiDAR scans. Beyond the Digital Surface Model (DSM or heightmap) provided by the height orthoimage, the proposed method cost-effectively generates a reflectance channel that is easily interpretable by human operators without relying on any optical acquisition, calibration and registration. Moreover, it commonly achieves very high resolutions (1cm ² per pixel), thanks to the typical sampling density of static or mobile LiDAR scans. Compared to orthoimages generated from aerial datasets, the proposed LiDAR orthoimages are acquired from the ground level and thus do not suffer occlusions from hovering objects (trees, tunnels and bridges), enabling their use in a number of urban applications such as road network monitoring and management, as well as precise mapping of the public space e.g. for accessibility applications or management of underground networks. Its generation and usability however faces two issues: (i) the inhomogeneous sampling density of LiDAR point clouds and (ii) the presence of masked areas (holes) behind occluders, which include, in a urban context, cars, tree trunks, poles or pedestrians (i) is addressed by first projecting the point cloud on a 2D-pixel grid so as to generate sparse and noisy reflectance and height images from which dense images estimated using a joint anisotropic diffusion of the height and reflectance channels. (ii) LiDAR shadow areas are detected by analyzing the diffusion results so that they can be inpainted using an examplar-based method, guided by an alignment prior. Results on real mobile and static acquisition data demonstrate the effectiveness of the proposed pipeline in generating a very high resolution LiDAR orthoimage of reflectance and height while filling holes of various sizes in a visually satisfying way.