Three-Dimensional Point Cloud Segmentation Using a Combination of RANSAC and Clustering Methods

There are challenges in performing 3D scene understanding on point clouds derived from drone images as these data are highly unstructured with no neighbouring information, highly redundant making the processing difficult and time-consuming and have variable density making it difficult to group and segment them. For proper scene understanding, these point clouds need to be segmented and classified into different groups representing similar characteristics. The approaches for segmentation differ based on the distinctiveness of each data product. Although newer machine learning-based approaches work well, they need large amounts of standardized labelled data which in turn require extensive resources and human intervention to obtain good results. Considering these, we have proposed a hybrid clustering-based hierarchical model for effective segmentation of dense 3D point cloud. We have applied the model to local data having a mix of man-made and natural vegetation with variable topography. The combination of RANSAC, DBSCAN and Euclidean method of cluster extraction proved to be useful for precise segmentation and classification of point clouds. The performance of the model has been assessed using Davies–Bouldin dbIndex-based intrinsic measures. The hybrid approach is able to segment 91% of the point clouds precisely compared to the conventional one-step clustering approach.