Open AccessFAST PROBABILISTIC FUSION OF 3D POINT CLOUDS VIA OCCUPANCY GRIDS FOR SCENE CLASSIFICATION
Author(s) -
Andrea A. Kühn,
Hai Huang,
Martin Drauschke,
H.J. Mayer
Publication year - 2016
Publication title -
isprs annals of the photogrammetry, remote sensing and spatial information sciences
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.356
H-Index - 38
eISSN - 2194-9042
pISSN - 2196-6346
DOI - 10.5194/isprsannals-iii-3-325-2016
Subject(s) - point cloud , computer science , artificial intelligence , segmentation , outlier , probabilistic logic , computer vision , range (aeronautics) , process (computing) , point (geometry) , noise (video) , scalability , octree , pattern recognition (psychology) , image (mathematics) , database , mathematics , materials science , geometry , composite material , operating system
High resolution consumer cameras on Unmanned Aerial Vehicles (UAVs) allow for cheap acquisition of highly detailed images, e.g., of urban regions. Via image registration by means of Structure from Motion (SfM) and Multi View Stereo (MVS) the automatic generation of huge amounts of 3D points with a relative accuracy in the centimeter range is possible. Applications such as semantic classification have a need for accurate 3D point clouds, but do not benefit from an extremely high resolution/density. In this paper, we, therefore, propose a fast fusion of high resolution 3D point clouds based on occupancy grids. The result is used for semantic classification. In contrast to state-of-the-art classification methods, we accept a certain percentage of outliers, arguing that they can be considered in the classification process when a per point belief is determined in the fusion process. To this end, we employ an octree-based fusion which allows for the derivation of outlier probabilities. The probabilities give a belief for every 3D point, which is essential for the semantic classification to consider measurement noise. For an example point cloud with half a billion 3D points (cf. Figure 1), we show that our method can reduce runtime as well as improve classification accuracy and offers high scalability for large datasets.