Automated camera parameters estimation and occlusion-aware visibility mapping of urban CCTV using asynchronously acquired MMS LiDAR

Pre-installed urban CCTV cameras form a dense, city-wide sensing network with significant potential for urban geospatial monitoring. However, most deployed units lack accurate camera parameters including position, orientation, and Field of View (FoV) geometry, which prevents their systematic integration into urban geospatial information systems. This study presents a fully automated pipeline that estimates these parameters and produces occlusion-aware visibility polygons as Geographic Information System (GIS) vector outputs from asynchronously acquired Mobile Mapping System (MMS) Light Detection and Ranging (LiDAR) point clouds. The pipeline needs no on-site access, no physical calibration targets, and no prior knowledge of the camera’s interior orientation, viewing direction, or lens distortion. The only positional input is an approximate horizontal camera location, such as that recorded in a municipal camera inventory or measured with consumer-grade GNSS. A coarse-to-fine strategy bridges the radiometric and temporal gaps between the LiDAR data and the optical imagery. Installation heights are inferred from the geometric signature of the pole in the point cloud, the approximate viewing direction is found by intensity-image matching on rendered panoramas, and precise 2D-3D correspondences are then established by deep feature matching with multi-stage spatial filtering. Camera parameters (interior orientation, exterior orientation, and radial distortion) are jointly estimated with a photogrammetric projection model that requires no initial approximations. An occlusion-aware visibility module then casts 3D rays against the full point cloud to separate ground that the camera can actually see from ground hidden by buildings and vegetation, and the boundaries are exported as GIS-ready vector polygons. Experiments on 21 urban CCTV cameras give a mean 3D camera position error of 0.85 m (σ = 0.33 m) despite a six-month gap between the MMS survey and the CCTV imagery. The resulting visibility polygons reach a mean Intersection over Union (IoU) of 0.721 against manually delineated ground truth. On the common 20-camera subset used for the manual baseline, the proposed method reduces the mean 3D position error by 81.2% (0.87 m versus 4.62 m) and is more accurate for 19 of the 20 cameras, supporting its reliability for large-scale deployment.