Rockfall detection from terrestrial lidar by clustering approach with non-parametric algorithm

This research aims to improve the rockfall detection framework from 3D point clouds by applying data analysis, statistics, and mathematic concepts. The general framework consists of three main steps: preprocessing, clutter removal, and rockfall isolation (using spatial clustering algorithm). In this study, we improve clutter removal and rockfall isolation processes, because the experiment following the typical framework found the two pain points. The first pain point is in the clutter removal process. Due to the suffering of high computational costs when data points increasing high.For the clutter removal process, two algorithms based on grid density were developed. The proposed algorithm can reduce the computational complexity inexpensively compared to the standard algorithm based on nearest neighbor distance (NNCR). Following the experimental results, the proposed method’s performance is comparable with the NNCR based algorithm result in terms of identifying rockfall events in terms of accuracy but it yields low recall values. The advantage of this method is speed, based on the simulation event with 52,000 data points, the proposed method is about 16 times faster.The algorithm was improved further by applying the grid-density-based algorithm recursively at multiple scales to improve a low recall problem due to point lost. Experimental results showed that the improved proposed method could recover data points lost in the previous algorithm. The proposed method can solve a low recall problem while still maintain good accuracy. Furthermore, it is even faster than the previously proposed method.For the rockfall isolation process, a non-parametric algorithm was proposed to solve the input parameter issue that the traditional DBSCAN, which requires two parameters. The proposed DC-DBSCAN algorithm can automatically customize the parameters. The simulation results showed that the proposed method is better than the traditional DBSCAN in the purity score. The purity score of the proposed method was 96.22%, while the conventional DBSCAN was 91.09%. In addition, the proposed algorithm can separate events better than the traditional DBSCAN.