Segmentation and Classification of 3D Urban Point Clouds: Comparison and Combination of Two Approaches

摘要: Segmentation and classification of 3D urban point clouds is a complex task, making it very difficult for any single method to overcome all the diverse challenges offered. This sometimes requires combination several techniques obtain desired results different applications. work presents compares two approaches segmenting classifying clouds. In first approach, detection, segmentation objects from clouds, converted into elevation images, are performed by using mathematical morphology. First, ground segmented detected as discontinuities on ground. Then, connected watershed approach. Finally, classified SVM (Support Vector Machine) with geometrical contextual features. The second employs super-voxel based approach in which cloud voxels then super-voxels. These clustered together an efficient link-chain form objects. local descriptors features basic object classes. Evaluated common dataset (real data), both these methods thoroughly compared three levels: classification. After analyses, simple strategies also presented combine methods, exploiting their complementary strengths weaknesses, improve overall results.