基于点云数据的自然地形分类算法

Nature Terrain Classification Using Point Cloud Data

提出一种自然地形分类算法以解决移动机器人导航中的环境理解问题。在单几何特征地形分类方法的基础上,提出的算法使用从点云中提取的复合特征训练分类器。复合特征向量包含一个点几何特征和颜色特征。算法首先计算点的坐标协方差矩阵和平均法向量协方差矩阵的特征值作为几何特征。然后通过标定激光雷达坐标系和相机坐标系使点得到颜色信息,把点的颜色作为颜色特征加入复合特征向量。算法使用最大期望-高斯混合模型(EM-GMM)训练一个分类器,训练数据由人工标注。实验结果表明,与单几何特征分类方法相比,复合特征分类方法在对自然地形分类时能得到更高的正确率。

A nature terrain classification algorithm is proposed to deal with the environment understanding problem in moving robot navigation.Based on the single geometrical feature terrain classification method,a complex feature extracted from point cloud is used to train a classifier.The complex feature vector includes both geometrical feature and color feature of a point.The algorithm computes the eigenvalue of the coordinate covariance matrix and the average normal vector of a point as a geometrical feature.The points obtain color information by matching coordinates of Lidar and camera.The color of points is added into complex feature vector as a color feature.An expectation-maximization Gaussian mixture model（EM-GMM） is employed to train a classifier.The training data are labeled by man.The experimental results show that： compared with the single geometric feature classification method,the complex feature classification method obtains higher correctness in classifying nature terrain.