Segmentation And Classification Of 3D Color Point Cloud Data

With the continuous development of 3D color point cloud data processing technology, the 3D color point cloud data has been widely concerned in the fields of robot, reverse engineering, laser remote sensing, virtual reality and human computer interaction. Segmentation and classification of 3D color point is a key step in many fields, especially in the field of mobile robot. It is the prerequisite and foundation for the complete autonomy.3D color point data by combining laser rangefinder and CCD camera, restore the original scene information more fully, but the three dimensional point cloud do not stay in the small scene, in the practical application of the number of data points in general about 100000. In this way, because of the complexity of the scene, the large number of noise points, and more interference, the correct segmentation and classification is a difficult task. The paper is mainly based on three-dimensional point cloud, supplemented by images. It is based on the geometric features, neighborhood features and color features of 3D point cloud to study the features of the scene.For 3D color point cloud data segmentation, this paper proposes a 3D color point cloud segmentation method based on principal component feature of the ball, using hierarchical thinking to segment 3D color point cloud. Firstly, the 3D color point cloud classification into point of point, linear point and surface point, not only to achieve the classified according to features of the 3D point cloud data value. Furthermore, it puts forward the classification method based on the point features, linear features and area features of 3D point cloud. Based on the construction features of main components of ball from the outside to the inside in turn is the normal vector of the surface of layer, the tangent vector of linear layer, layer of the second eigenvector, then opposite point position segmentation, to prepare for the extraction of ground. Meanshift clustering algorithm is used to extract the ground. Finally, unground points are clustered by DBSCAN algorithm based on density. The method to achieve accurate ground extraction, and can ensure that the system can be extracted ground, and the use of DBSCAN for 3D color point cloud data segmentation based on the RGB value of color of 3D point cloud data.For the classification of 3D color point cloud data, this paper proposes a 3D color point cloud data classification method based on conditional random field model. The model constructs three models respectively:point model, two order model and high order clique model. The two order model is the edge of two adjacent points. High order clique model consists of points in the neighborhood, and is the use of 3D color point cloud based on color segmentation result. On the basis of this model, the features of point model, second order model and high order clique are extracted, and the feature of the high order clique of whole regiment in all points. Features and the training sample label together constitute the CRF model of three dimensional color point cloud data. Using subgradient iterative learning method, and it concluded that the most appropriate model parameters are obtained, which are inferred labels. Experimental results show that this method has higher classification accuracy and recall rate.