Outdoor Scene Segmentation And Classification Based On 3D Laser Point Clouds

3D outdoor scene segmentation and classification are two important tasks in robotics and computer vision. Recently, a variety of approaches have been proposed to handle these problems, but they are still challenging because of the complexity of the scene and the great number of 3D points in the scene. The goals of this paper are to present a fast and accurate segmentation approach for 3D point cloud scene and to perform a contextual classification using the segmentation result.In this paper, a fast and accurate 3D scene segmentation approach based on region growing algorithm is proposed. The total number of segmented patches in this algorithm depends on the complexity of actual 3D scene instead of being artificially set. Thus, the segmentation result is accurate and reasonable. Since no iteration procedure for each point is used in this approach, it can perform fast segmentation than traditional segmentation algorithms like k-means based segmentation. Before actual segmentation, a matrix of 3D point was constructed from sequential 2D laser data and all 3D points are included in the matrix. The significant meaning of the matrix is that all adjacent relations between 3D points can be easily found in it. The remaining procedure can be summarized as:(1) Estimate the normal vector of each point using its two neighboring points; (2) Give a plane/not plane label to each point based on its normal vector and its neighbors'normal vectors; (3) Extract ground points based on the height and normal vector direction of the points; (4) Generate segmented patches using region growing algorithm.Contextual classifiers have proven to conduct a better performance than traditional local classifiers in the application of point cloud classification. As a popular discriminative model, a Conditional Random Field (CRF) model is extensively used in contextual classification of computer vision and robotic. Thus, it is utilized in the scene classification framework. When pairwise interaction is used, the inference will take large amount of time because of the highly linked random field. In order to reduce computation time, this paper construct nodes and cliques of the random field without constructing edges of the random field. The segmented patches are selected as CRF nodes and the high-order cliques are generated by a clustering algorithm which is similar to the above segmentation algorithm. To speed up the classification, a few but crucial features are calculated to describe nodes and high-order cliques. Based on four data sets, a systematic experiment is carried out to verify the segmentation and classification method, and the experimental results show the effectiveness of the proposed methods.