An Image Dataset for Illegal Road Crossing Detection in Bangladesh

Description

This dataset supports research on pedestrian detection and illegal road-crossing analysis in crowded urban settings. The images were collected in Dhaka, Bangladesh, from locations including the University of Dhaka campus, Elephant Road, Bakshibazar, Motijheel, Gulistan, and the Dhaka-Chittagong Highway. The dataset contains variations in viewing angle, lighting conditions, pedestrian scale, traffic density, occlusion, and background complexity. The package contains original JPG images along with YOLO-format object-detection datasets created for YOLOv8 and YOLO11 training experiments. Two labeling schemes are included. The first is a three-class road-position scheme consisting of person_middle, person_side, and person_sidewalk. The second is a single-class crossing scheme in which all objects are labeled as person_middle. Annotations consist of manually created, axis-aligned bounding boxes in normalized YOLO format. The current workspace contains multiple exported variants under the YOLO_Dataset/ directory, including both multi-class and single-class datasets at 640, 960, and 1280 image resolutions. The dataset is organized into an Original_Images/ directory and a YOLO_Dataset/ directory. The Original_Images/ directory contains the source JPG photographs (total 694 images ). The YOLO_Dataset/ directory contains the following dataset variants: multi_class_road_crossing_640_yolov8 multi_class_road_crossing_960_yolov8 multi_class_road_crossing_960_yolov11 multi_class_road_crossing_1280_yolov8 single_class_road_crossing_640_yolov11 single_class_road_crossing_960_yolov8 single_class_road_crossing_960_yolov11 single_class_road_crossing_1280_yolov11 Each dataset folder contains train, valid, and test splits. Each split includes images/ and labels/ directories, while each dataset folder also contains a data.yaml file. Annotation Classes: For the multi-class datasets, the class mapping is: 0 = person_middle 1 = person_side 2 = person_sidewalk For the single-class datasets, the class mapping is: 0 = person_middle

Steps to reproduce

The YAML files included in the package are consistent with their respective annotation schemes. Multi-class datasets use three classes: person_middle, person_side, and person_sidewalk. Single-class datasets use one class: person_middle.

Institutions

• University of Dhaka

  Dhaka Division, Dhaka

Categories

Licence