Single Stem Extraction from Ground Laser Point Cloud Based on Point Classification and Improved Mean Shift Clustering
In view of the problem of over-segmentation and under-segmentation generated by the difficulties of individual tree recognition in a large-scale scene caused by the shadow-ing effect of understory and dense distribution of crown,a stem detection method based on point cloud classification and mean shift clustering is proposed in this paper,which can pro-vide ground reference data for forest inventory and enrich the methods of the stem detection. Firstly,the ground points are separated by Cloth Simulation Filter (CSF) for the ground model establishment. Then,the point cloud features for non-ground points are constructed by adaptive KNN and the stem points based on Random Forest (RF)are extracted. Finally,an improved method based on Mean Shift clustering is used for single stem extraction. The non-stem clusters are removed by adaptive filtering and the stem detection can be realized by the RANSAC cylinder fitting from the sliced point cloud. The results show that the precision of the two research plots based on the proposed method are 90.06％ and 91.00％ and the recall rates are 93.33％ and 91.46％ respectively. In addi-tion,the validity and accuracy of mean shift clustering in sin-gle stem extraction are verified and analyzed by single tree clustering experiment.
terrestrial laser scanning (TLS)RFmean shift clusteringRANSAC cylinder fittingindividual tree position
万方数据
维普
