三次元点群への円柱あてはめによる樹幹自動検出

抄録

Quantitative assessment of forest resources is an important topic from the point of view of the absorption of greenhouse gases. This article presents two contrasting approaches to automatic detection of tree-trunks by fitting right cylinder models to 3D point clouds taken in real forest land as well as those synthesized for analysis purposes. Volumetric index of tree-trunks can be estimated by those model parameters. The first approach segments point cloud into overlapping local bounding boxes (windows) and cylinder models and terrain models are calculated in each window by nonlinear and linear least-square methods. This method secures the detection accuracy of tree-trunks by means of the rejection of false-positives through outlier detection and the avoidance of false-negatives through redundancy of search space. The second approach formulates the present problem as an engergy minimization and detects multiple cylinder models concurrently by a graph-cuts algorithm. Experimental results provide enough evidence of effectiveness of the proposed approaches.