Learning to Expect the Unexpected: Benchmarking and Detecting Unexpected Road Hazards

Unexpected road hazards remain a critical challenge for driving safety, as they exhibit irregular appearances, vary widely in scale, and frequently arise under adverse environmental conditions. Existing detection approaches often struggle to achieve reliable localization under challenging conditions such as rain, fog, low-light environments, and small-hazard scenarios. To address these limitations, we propose ROAD-DETR, a difficulty-aware, transformer-based detection framework specifically designed for robust detection of unexpected road hazards. Specifically, it enhances a transformer encoder–decoder architecture by introducing three task-specific modules: (1) a Hazard-Aware Multi-Scale Fusion Module for multi-scale feature aggregation, (2) a Difficulty-Aware Feature Enhancement Module to enhance visually degraded hazard representations, and (3) a Hazard Query Refinement Module for adaptive refinement of transformer object queries. Training and evaluating such a model is challenging because most existing road-scene datasets focus on common traffic participants, while hazards such as Accident Vehicles, Fallen Trees, Icy Roads, Oil Spills, Road Debris, and Surface Water remain underrepresented. In response, we introduce ROAD-6-Det, a benchmark for unexpected road hazard detection with bounding-box annotations. Extensive experiments on ROAD-6-Det using a wide range of state-of-the-art YOLO-based and transformer-based detectors show that our ROAD-DETR achieves 94.2% mAP@50, 72.1% mAP@50:95, 93.2% precision, 91.1% recall, and an F1-score of 93.5%, outperforming existing models. The proposed ROAD-DETR framework and the ROAD-6-Det benchmark provide a strong foundation for future research on unexpected road hazard detection. For reproducibility, the code and dataset will be released publicly.