Investigating the Impact of Training Data Set Selection for Pole-Like Object Classification in Mobile Laser Scanning Data

The presence and functionality of traffic signs, traffic lights, and streetlamps are vital for ensuring road user safety and efficient traffic management. These devices typically are mounted on pole-like structures, and their proper placement impacts their effectiveness. Traditional methods of surveying these streetside assets are labor-intensive, time-consuming, and costly. Advanced three-dimensional (3D) reality capture technologies, such as mobile laser scanning (MLS) provide a promising alternative. This study enhances transportation asset management by leveraging MLS data for the classification of pole-like objects and surrounding features using machine learning (ML) models, including random forest and neural networks. The research took the survey-grade MLS data sets from the Oregon DOT (ODOT) across various urban and rural settings as an example to train and test different models, then analyzed the impact of the training data selection. The proposed experiment specifically focused on classification of pole-like objects in the point cloud data, and further tested the deep learning algorithm PointNet++ to assess its feasibility and performance in a realistic scenario. The models trained on diverse and well-represented data sets were found to perform better, whereas those trained on data sets with more unclassified and underrepresented objects struggled with generalization. Although random forest had higher accuracy in the experiment, neural networks were more consistent in terms of their performance across different data sets, which indicated that they would be more effective in real-world applications with limited object labels.