Segmentation-based classification for 3D urban point clouds

In this paper, we propose a unified classification framework for 3D urban point clouds. First of all, an efficient segmentation approach is utilized to segment 3D point clouds. For comparison, we employed two recently developed point clouds segmentation approaches. The first one is a region-growing-based segmentation algorithm by using robust saliency features and another one is a hierarchical-clustering-based algorithm. The former will produce large and complete segments while the latter will generate more small pieces of segments. Secondly, we extract a set of features from each segment for training and testing by using an effective classifier. For comparison, we employed two popularly used classifiers, Naïve Bayes (NB) classifier and support vector machine (SVM). Thirdly, the contextual constraints among objects is used to further refine the classification results based on segments via graph cuts optimization. Experimental results on 3D urban point clouds acquired by a vehicle LiDAR system illustrate that our proposed classification framework is effective.