Hydrant Segmentation and Extraction Using Deep Learning Models for Large-Scale Urban Point Clouds

This study addresses the gap in research concerning the segmentation of smaller infrastructure, such as fire hydrants, within large-scale point cloud models. The RandLA-Net model, an efficient and robust deep learning algorithm known for its ability to process vast quantities of data while maintaining high accuracy rapidly, was evaluated for this task. A comprehensive dataset was constructed using LiDAR technology to gather point cloud data from diverse urban scenes. The data was manually segmented to assign unique class identifiers to distinct objects within the point cloud. The model was trained using these annotated data, and its performance was assessed using Intersection over Union (IoU), a typical evaluation metric for segmentation tasks. The results demonstrated the model's impressive capability in segmenting different classes within large-scale urban point cloud data. However, segmenting smaller, less distinctive objects like fire hydrants revealed room for improvement. Future work should focus on improving the recognition of complex or underrepresented classes and enhancing training stability. The study's findings provide critical insights into the capabilities and limitations of the RandLA-Net model in the context of large-scale, real-world point cloud data segmentation tasks.