SPARSE POINT CLOUD FILTERING BASED on COVARIANCE FEATURES

This work presents an extended photogrammetric pipeline aimed to improve 3D reconstruction results. Standard photogrammetric pipelines can produce noisy 3D data, especially when images are acquired with various sensors featuring different properties. In this paper, we propose an automatic filtering procedure based on some geometric features computed on the sparse point cloud created within the bundle adjustment phase. Bad 3D tie points and outliers are detected and removed, relying on micro and macro-clusters analyses. Clusters are built according to the prevalent dimensionality class (1D, 2D, 3D) assigned to low-entropy points, and corresponding to the main linear, planar o scatter local behaviour of the point cloud. While the macro-clusters analysis removes smallsized clusters and high-entropy points, in the micro-clusters investigation covariance features are used to verify the inner coherence of each point to the assigned class. Results on heritage scenarios are presented and discussed.