AccidentDetection-9k: An Annotated Road Accident Image Dataset for Object Detection

This dataset contains approximately 9,000 annotated road-accident images developed for object-detection research. The images were collected through systematic, keyword-based Google image searches using an automated image-downloading tool. The dataset includes a wide range of visually identifiable accident scenarios, such as vehicle-to-vehicle collisions, vehicle rollovers, overturned vehicles, heavily damaged vehicles, and other road-crash scenes. These variations were included to represent different accident conditions, vehicle positions, viewing angles, backgrounds, and levels of damage. After collection, the images were uploaded to the Roboflow platform for annotation. Each accident region was manually labelled using a bounding box under a single object class named “Accident.” The annotations were reviewed and refined to reduce missing labels, inaccurate bounding-box boundaries, and inconsistencies in the labelling process. The final dataset was exported in YOLO annotation format and divided into separate training, validation, and testing subsets. Each subset contains an images folder and a corresponding labels folder. A unified data.yaml file provides the dataset paths and class information required for YOLO-based model training. This dataset can be used to develop, train, validate, and evaluate deep-learning-based road-accident detection models. It may also support research in traffic surveillance, intelligent transportation systems, road-safety monitoring, and automated emergency-response applications.