EdgeVision Dataset

The EdgeVision Dataset is a curated image dataset for helmet detection in motorcycle traffic scenes. It contains 2,392 high-resolution images annotated with bounding boxes across three classes: BikeWithRider, NoHelmet, and Helmet. Images were collected from real-world motorcycle traffic scenarios and annotated by five annotators. The dataset contains 8,275 bounding box annotations with an average image resolution of 1,154x1,411 pixels. This dataset was created as part of a broader study on federated learning-based helmet detection for intelligent transportation systems. Annotations are provided in three formats: Pascal VOC XML (one .xml per image, absolute pixel coordinates; class given by name), YOLO (one .txt per image, normalized coordinates), and COCO (single annotations.json). The class index order is fixed across all formats as BikeWithRider (0), NoHelmet (1), Helmet (2); YOLO integer labels use this index directly (see classes.txt), while COCO category_id equals the index plus one (1 = BikeWithRider, 2 = NoHelmet, 3 = Helmet). Suggested use cases include object detection model training and evaluation, motorcycle safety research and federated learning benchmarking.