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

The "Augmented Bangladesh Traffic Dataset for Deep Learning-based Vehicle Detection under Diverse Weather Conditions" was developed to support research on vehicle detection in heterogeneous Bangladeshi traffic environments under varying weather and lighting conditions.

The dataset contains 8,605 annotated traffic images representing real-world road traffic conditions in Bangladesh. The original dataset consisted of 4,762 images collected directly by the authors from multiple urban road intersections in Dhaka, Bangladesh, between March 2025 and February 2026.

Images were captured from several locations including Maghbazar, Mirpur-10, Kakrail, Gulshan, Dhanmondi, Shantinagar, Uttara, Mugda, and Asad Gate. All images were collected from publicly accessible traffic environments to ensure realistic urban traffic diversity.

To improve robustness against environmental variability and class imbalance, offline data augmentation techniques were applied exclusively to the training subset of the dataset. These techniques include horizontal flipping, rotation, brightness adjustment, simulated fog, simulated rain, and low-light enhancement. After augmentation, the total dataset size increased to 8,605 images.

All images were spatially normalized to a fixed resolution of 640 × 640 pixels using resizing and zero-padding while preserving the original aspect ratio.

The dataset includes bounding box annotations for 10 vehicle categories: Car, Bus, Truck, Bike, Rickshaw, Van, Bicycle, Leguna, CNG, and Emergency Vehicle (including Ambulance and Police vehicles). Annotation was performed using the Computer Vision Annotation Tool (CVAT) and exported in YOLO Detection format.

Additionally, each image includes global environmental metadata tags describing weather and illumination conditions such as Sunny, Rain, Fog, Night, Rainy-Night, Foggy-Night, High Light, and Low Light.

The dataset is partitioned into training (7,183 images), validation (949 images), and testing (473 images) subsets. The augmented images are included only within the training subset to improve model generalization under diverse weather conditions.

This dataset is intended for research on deep learning-based vehicle detection, intelligent traffic monitoring, smart traffic systems, and environment-aware computer vision applications for complex traffic environments.

Citation

If you use this dataset, please cite:

Mondal, Madhure Rani; Halder, Purbasha; Rahman, Anika (2026), “Augmented Bangladesh Traffic Dataset for Deep Learning-based Vehicle Detection under Diverse Weather Conditions”, Mendeley Data, V2, doi: 10.17632/2f9jp8bj45.2