Traffic Rule Violation Detection Dataset in Dhaka Urban Traffic Environment

It provides bounding-box annotations for two major motorcycle violations: helmet non-compliance (no helmet) and triple riding (three or more occupants on one motorcycle). The raw dataset contains approximately 1,100 manually annotated street traffic images.

Through offline geometric and environmental augmentation, including flips, brightness and exposure variation, Gaussian blur, and sensor noise, the dataset was expanded to 4,839 frames containing 30,506 bounding-box instances. Dataset Structure The repository contains two dataset variants and a metadata file: Raw_Dataset/ – Original, un-augmented images and annotations divided into standard splits: train/ – Training images with MS-COCO JSON annotations (_annotations.coco.json). valid/ – Validation images for model development and hyperparameter tuning. test/ – Held-out evaluation set containing 78 frames and 497 annotated instances. augmented_dataset/ – Offline-augmented images: Training split: 4,644 frames containing 29,250 bounding boxes.