Safety Driving-oriented Classification Based on Point Cloud for Environment Perception
K-Y. Du, J-H. Lyu, Y. Zhou, S-F. Wang and J-H. Yang
There has been a growing trend towards environment perception study of unmanned vehicles and the majority of approaches have adopted multiple sensor data fusion in order, which increases the complexity of the experiments. Proposed herein is a method to recognize the objects around the unmanned vehicle with support vector machine (SVM) classifier that only relies on point cloud data. For the purposes of reducing time consuming, the study employs the cube to segment the interest area. The cube presents a spatial domain with specific edge length that contains several certain properties. A frame of processed point cloud was segmented into a series of cubes arranged tightly in line with the coordinate information. Every single cube was regarded as a processing unit to extract features like raw intensity, height, distance, entropy, the Eigenvalues of the covariance matrix of a three-dimensional (3-D) coordinate matrix in an ascending order and normal vector. By means of the principal component analysis (PCA) we chose the appropriate dimensionality to train model. The results demonstrated that the method could achieve the aim of only using point cloud data for environment perception with a prominent classification accuracy of the 97.54% average.
Keywords: LiDAR, driving, safety, support vector machine (SVM), principal component analysis (PCA), point cloud, classifier, environment perception