ScholarWorks@성균관대학교

초록

Detecting unexpected road hazards is essential in ensuring safe and adaptive navigation of autonomous vehicles (AVs). However, existing datasets are significantly limited in unexpected road object categories, fail to capture the complexity of real-world scenarios, and lack risk-level categorization. Unexpected objects can appear differently in real-world traffic scenarios and, therefore, pose various kinds of risks for AVs. This requires a robust model to recognize the objects and control the AV navigation. The lack of diverse datasets hinders the development of such models. To address these challenges, we present a new dataset called ROAD-6, composed of six classes, including ''Accident Vehicles'', ''Fallen Trees'', ''Oil Spills'', ''Icy Road Patches'', ''Road Debris'', and ''Surface Water'', which are categorized into three risk levels (Low, Moderate, and High). ROAD-6 reflects complex real-world scenarios that are suitable for model generalization. We benchmark state-of-the-art recognition models and propose a framework leveraging MobileViT, achieving an accuracy of 93.15%, to recognize unexpected hazards effectively. ROAD-6 lays the groundwork for risk-aware frameworks, contributing to safer and more reliable autonomous navigation. The dataset and code are publicly available at GitHub. © 2025 ACM.

키워드