Utilizing deep learning models and LiDAR data for automated semantic segmentation of infrastructure on multilane rural highways

This paper presents two Transformer-based approaches for automating the extraction of rural multilane highway infrastructure elements from light detection and ranging data. The first approach uses the Point Transformer v2 model with four additional attributes as input, while the second adapts self-attention and cross-attention mechanisms for point-wise classification. Experiments on 2.5 km of highway in Alberta, Canada, demonstrate the effectiveness of both methods. The first approach achieved a mean Intersection over Union (IoU) score of 78.29% and a mean F1 score of 86.48%, with most class accuracies exceeding 95%. The second method achieved a mean IoU score of 86.03% and a mean F1 score of 92.21%. This research advances automated infrastructure extraction techniques, providing transportation agencies with efficient inventory methods for rural highway infrastructure. The study has implications for autonomous driving, crash environment reproduction, highway safety understanding, big data analysis, maintenance planning, and asset management, highlighting its relevance and importance in modern transportation systems.