Automated lamp post extraction using drone based photogrammetry

Abstract

Utility poles, particularly lamp posts, are essential for urban planning, infrastructure management, and road safety. Accurate roadside asset inventories are increasingly required to support maintenance and safety assessments. UAV photogrammetry has emerged as a potentially cost-effective and efficient approach for infrastructure mapping due to its rapid data acquisition and high spatial resolution. However, extracting pole-specific information from UAV-derived point clouds remains challenging and labour-intensive, especially in newly developed areas. While Light Detection and Ranging (LiDAR) has been widely applied for pole detection, limited studies have explored UAV photogrammetry as a standalone solution. This study proposes a proof-of-concept automated workflow for extracting lamp posts from UAV-derived point clouds using clustering-based segmentation techniques. The method integrates planar grid filtering for feature isolation and K-means clustering for object identification, with preprocessing steps including voxelisation and statistical outlier removal. The workflow was implemented using geometrically defined parameters, which are expressed in metres in accordance with the coordinate system of the dataset. Parameter values, including voxel size (0.1 m), neighbour distance (10 m), grid resolution (0.1 m), standard deviation (2.0), and lower-bound height (8 m), were determined through empirical evaluation. The method was applied to a dataset of 402 UAV images to generate a dense point cloud. Clustering analysis identified six distinct objects, consistent with the number of lamp posts observed in the study area. The extracted heights showed a mean difference of 0.02 m when compared with field measurements. The results demonstrate the potential feasibility of the proposed approach for automated extraction of pole-like structures from UAV photogrammetric point clouds under the tested controlled conditions. However, the validation is limited to a single study area and should be interpreted as a proof-of-concept. Further evaluation using larger and more diverse datasets is required to assess robustness, scalability, and applicability across different environments.