Unseen Road Type Detection in Road Networks for Intelligent Transportation Systems

Existing methods for road type (highway, trunk road, etc.) classification assume that every road in real environments belongs to a road type seen during training. However, there are unseen road types in real-world scenarios. Thus, unreliable classification of an unseen-type road into a seen road type can cause critical safety issues in road-related applications. In this paper, we introduce a new framework to detect unseen road types. To this end, we adopt an out-of-distribution (OOD) detection approach studied in the deep learning field. However, conventional graph-based node-level OOD detection methods cannot be directly applied to the unseen road type detection problem since roads are represented by edges in road networks. To resolve this problem, we establish a new formulation of edge-level OOD detection and propose a novel energy propagation scheme on a line graph transformed from a road network to obtain OOD scores. Experimental results on real-world road networks demonstrate the effectiveness of our method, achieving state-of-the-art performance in unseen road type detection.