Augmented Bangladesh Traffic Dataset for Deep Learning-based Vehicle Detection under Diverse Weather Conditions

Description

The "Augmented Bangladesh Traffic Dataset for Deep Learning-based Vehicle Detection under Diverse Weather Conditions" (ABTD) supports research on vehicle detection in heterogeneous Bangladeshi traffic under varying weather and lighting conditions. The dataset contains 8,605 annotated traffic images. The original set consisted of 4,772 images collected directly by the authors from nine road intersections in Dhaka, Bangladesh, between March 2025 and February 2026: Maghbazar, Mirpur-10, Kakrail, Gulshan, Dhanmondi, Shantinagar, Uttara, Mugda, and Asad Gate. All images were captured by the authors from publicly accessible traffic environments; none were obtained from newspapers, social media, or third-party sources. To balance the representation of environmental conditions, offline augmentation was applied only to the training subset, at the condition level: each original training image was replicated by condition rarity (Foggy-Night ×15, Rainy-Night ×6, Night ×2, Foggy-Day ×1, Rainy-Day ×1, Sunny-Day ×0) using horizontal flipping, rotation, brightness/contrast adjustment, RandomFog, and RandomRain. This expanded the training subset from 3,350 to 7,183 images. All images were resized to 640 × 640 pixels; the original aspect ratio is not preserved. Formats: JPG (4,556), PNG (3,993), JPEG (56). The dataset includes bounding-box annotations for 10 vehicle categories: Car, Bus, Truck, Bike, Rickshaw, Van, Bicycle, Leguna, CNG, and Emergency Vehicle (Ambulance and Police). Annotation used CVAT and was exported in YOLO format, with each label file containing only the ten vehicle classes (indices 0–9), as defined in data.yaml. Weather and illumination are provided as image-level metadata in metadata.csv, with columns: image, split, weather (Sunny, Rain, Fog, Clear), time_of_day (Day, Night), illumination (High_Light, Low_Light), augmented, n_objects. Of 8,605 tags, 6,991 were assigned during annotation and 1,614 recovered from capture filenames. The dataset is partitioned into training (7,183), validation (949), and testing (473); augmented images appear only in training. It contains 39,488 annotated instances. A YOLOv8n baseline attains a test mAP@0.5 of 0.852. Version 3 corrects the Version 2 annotation format, in which the seven environmental conditions had been exported as YOLO class indices 0–6 with full-image boxes alongside the vehicle classes. Version 3 removes these boxes, renumbers the ten vehicle classes to 0–9, and moves the environmental tags to metadata.csv. Citation: Mondal, Madhure Rani; Halder, Purbasha; Rahman, Anika (2026), “Augmented Bangladesh Traffic Dataset for Deep Learning-based Vehicle Detection under Diverse Weather Conditions”, Mendeley Data, V3, doi: 10.17632/2f9jp8bj45.3

Institutions

• Stamford University Bangladesh

  Dhaka Division, Dhaka

Categories

Licence