Coarse-to-Fine Classification of Road Infrastructure Elements from Mobile Point Clouds Using Symmetric Ensemble Point...

Area of Science:

• * Geospatial data analysis
• * Computer vision
• * Machine learning for infrastructure classification

Background:

• * Accurate classification of road infrastructure from mobile laser scanning (MLS) data is crucial for road asset management and autonomous vehicle navigation.
• * Deep learning models offer generalizability but can produce false predictions, while traditional clustering methods may lack adaptability.
• * Existing methods face challenges in accurately identifying diverse road elements in complex environments.

Purpose of the Study:

• * To develop and evaluate a novel hybrid method for enhanced classification of road infrastructure point clouds.
• * To improve the accuracy and reliability of road asset identification using mobile laser scanning data.
• * To provide a robust solution for practical applications in transportation network management.

Main Methods:

• * A novel approach combining a Symmetric Ensemble Point (SEP) network for coarse classification with Euclidean Cluster Extraction (ECE) for refinement.
• * The SEP network utilizes symmetric functions for multi-scale feature extraction and ensemble methods for optimal sub-sampling.
• * ECE algorithm is employed to correct misclassified points from the initial SEP network output.

Main Results:

• * The proposed SEP-ECE method successfully classifies six key road infrastructure elements: road surfaces, buildings, walls, traffic signs, trees, and streetlights.
• * Achieved an average classification accuracy of approximately 99.74%.
• * Demonstrated a 3.97% improvement in overall accuracy compared to the PointNet baseline.

Conclusions:

• * The SEP-ECE method offers a highly accurate and reliable approach for road infrastructure classification from MLS data.
• * The high accuracy achieved makes this method suitable for practical deployment in transportation asset management and autonomous driving systems.
• * This hybrid approach effectively addresses limitations of purely deep learning or traditional clustering methods.