3D Point Cloud Classification Based on Discrete Conditional Random Field

Abstract

The advantage of the conditional random field (CRF) lies in the construction of the discriminant model and efficient parameter optimization. In the topic of the classification of three-dimensional point cloud, the parameters of the CRF are usually learnt through Gradient Descent and Belief Propagation, in order to optimize the objective energy function of the CRF. These optimization methods do not guarantee the highest global classification accuracy with a high classification accuracy on the smaller classes. In addition, differential features of the point cloud are not sufficiently utilized. In this paper, we use the local geometric shape features to construct the CRF, including the nearest neighbor tetrahedral volume, Gaussian curvature, the neighbourhood normal vector consistency and the neighbourhood minimum principal curvature direction consistency. We propose four discrete criteria for CRF parameter optimization to design the explicit functions, and present concrete solution procedures, in which Monte Carlo method and supervised learning method are employed to estimate the CRF parameters iteratively. Experimental results show that our method can be applied to the classification of the scene of 3D point cloud with plants, especially under the proposed second criterion that maximizing the accuracy with interclass weights. It can be used to improve significantly the classification accuracy of small scale point sets when different classes have great disparity in number.