Research On Terrain Classification Method Based On 3D Color Point Cloud

Through calibration of camera and lidar, the three-dimensional point cloud data will obtain the color information and become the colored three-dimensional point cloud data. The colored point cloud data is closer to the real world and can reflect the environmental information better for a robot. Therefore, it is significant for robots’ environmental understanding to study how to classify a natural terrain into several categories based on colored point cloud data.This paper focus on studying the terrain classification methods based on colored three-dimensional point cloud data. The contents of the study include colored point cloud data preprocessing, point cloud data robust feature extraction, variable-scale three-dimensional grid map establishing and terrain classification based on variable-scale three-dimensional grid map. The main research results of this paper are as follows:In order to make the scattered point cloud data with a topological relationship, a displacement calculation method is used to calculate the Morton code for the point cloud data. To reduce the impact of noise on point cloud data feature extraction and classification, the neighborhood average algorithm is used for point cloud data denoising. Experimental results show that according to the Morton code, each point can easily access its k neighbors. Most of the noise in raw data can be detected after denoising.In order to reduce the impact of outliers on point cloud data local feature extraction, an outlier detection method based on local fitted plane and RANSAC is proposed. First select the local best fitted plane by using the RANSAC. Then detect the outliers by calculating the Mahalanobis distance of each point to the fitted plane. Experimental results show that this method can improve the robustness of feature extraction.In order to improve the classification result, the impact of different feature extraction methods on natural terrain point cloud data classification is studied. Experimental results show that the accuracy of the classification results can be effectively improved by combining the color feature and geometric features as the category feature.According to the distribution characteristics of lidar collection points, dense in the vicinity and sparse in the distance, a terrain classification method based on variable-scale three-dimensional grid map is proposed to classify an unknown terrain into four categories, which includes roads, lawns, buildings and trees. First, establish a variable-scale three-dimensional grid map. Then use the robust point cloud feature extraction methods to extract the features of voxels. Finally, employ the classifiers based on TWSVM to classify the voxels into several categories. Experimental results show that the method can reduce the number of voxels while ensuring accuracy, reduce the noise and have good classification results on real data sets.According to the characteristics of the voxels, large amount and non-uniform distribution, by taking the idea of image processing, in which by reusing the previously computed results to improve the efficiency, an optimization method which could self-adapt to the change of voxel density to improve the efficiency in the process of calculating is proposed. Experiments show that it can effectively improve the computing speed when calculate the three-dimensional grid map.