Comparison of point cloud filtering methods with data acquired by photogrammetric method and RGB-D sensors

Abstract

Point clouds (PCs) are inevitable sources to generate digital solid model-based applications such as reverse engineering, differential 3D modelling, 3D sensing and modelling of environments, scene reconstruction, augmented reality. Photogrammetric methods, Terrestrial Laser Scanners and RGB-D sensors are relatively common among the technologies used to capture PCs. Because of their structural characteristics, measuring systems produce large amounts of noise that cannot be precisely predicted in type and amplitude. Due to the noisy measurements, the spatial orientations of the differential surface particles and the spatial locations of the corner points have a certain degree of deformation. In order to increase visual, spatial and physical quality of the solid model, which is frequently used in reverse engineering, PCs must be filtered to discard noise and outlier. In this paper PC produced from different methods was filtering with Shepard Inverse Distance Weighting method, Gaussian Filtering method, Single Value Decomposition Based Plane Fitting method and Optimization Based Plane Fitting method. Backtracking Search Optimization Algorithm (BSA) was used to fitting plane. Experimental results were compared visually and statistical according to the number of neighborhoods. The results showed that Backtracking Search Optimization based filtering supplied better noise smoothing results than its competitors.