Towards Real-Time Drone Vision for Road Safety

Traffic accident detection is a critical challenge for emergency response systems, where timely situational awareness can be the difference between life and death. This paper presents preliminary results of the VISTA project—a collaboration between the University of Bari Aldo Moro and the Italian Ministry of Infrastructure and Transport—which aims to equip first responders with real-time aerial intelligence by dispatching AI-powered UAVs immediately upon accident alert. As an initial phase of this broader initiative, we introduce CrashVista, a specialized dataset of 1,000 annotated aerial and elevated-view images covering diverse accident types, viewpoints, and environmental conditions. Given the limited scale of available accident imagery, we evaluate a broad set of models under two complementary paradigms—zero-shot inference and supervised domain adaptation—to quantify the extent to which the annotation effort provides a tangible advantage over off-the-shelf open-vocabulary models. Evaluated approaches span open-vocabulary detectors, multimodal language models, segmentation-based methods, and multiple YOLO families across different scales. Together, these experiments establish a benchmark for the VISTA system and demonstrate the feasibility of vision-based automated accident detection as a basis for future work on severity assessment, casualty localization, and real-time emergency guidance. The code and the dataset are available at https://github.com/pasqualedem/VISTA.