Automatic segmentation of plant point cloud from Multi-view stereo

摘要

In this paper, a method for automatic segmentation of plant point cloud is proposed. We get the quasi-dense point cloud of plant from Multi-view stereo reconstruction based on surface expansion. The Adaptive Normalized Cross-Correlation algorithm is used as matching cost to match points of interest in two images, which is robust to radiometric factors and can reduce the fattening effect of boundaries. An efficient segmentation framework is proposed to segment plant from background. After oversegmenting the input point cloud, we extract the 3D feature for each segment and calculate conditional label probabilities using a Random Forest classifier. The out of the classifier is to initialize the unary potentials of a dense CRF whose optimization yields the final labeling. A highly efficient approximate inference algorithm based on mean field approximation is applied to the dense CRF models, in which the pairwise edge potentials are defined by Gaussian kernel. Experimental results show that our segmentation framework based on dense CRF can separate plant from background effectively.