Multiscale point feature object localization for hydrant surveying using LiDAR

Object localization in outdoor point clouds is essential for urban scene understanding in numerous applications, especially in land surveying. The advent of terrestrial laser scanning (TLS) LiDAR and Deep Learning methods can re-duce the time surveying urban objects in real-world situations. This paper proposes an automatic and effective ob-ject detection and key-point feature detection pipeline for surveying hydrant objects on dense point-cloud scenes. The proposed method consists of two stages. In the first stage, a multiscale voxelization strategy is proposed to reduce the computational load and complexity of dense point clouds, then hierarchical features are extracted to localize the in-terest object. In the second stage, we introduced an auto-matic strategy to seek the hydrant's centroid point using a learning-based method with KD-trees. We exploit the features using PointNet++ and compare the performance under different configurations in both stages. The proposed pipeline demonstrates robustness under challenging sce-narios and represents an intuitive solution for surveying ur-ban objects in dense point-cloud data.