BDRoad-Sense: A Benchmark Dataset for Road Surface Detection and Safety Enhancement

The BDRoad-Sense dataset was developed to support research on automated road surface monitoring and road safety assessment in real-world environments. It is a multi-class image dataset that represents common road surface conditions and infrastructure components that can directly affect driving safety and transportation planning. The dataset includes five annotated categories: Major Damage, Minor Damage, Normal Road, Manhole, and Speed Breaker, allowing both damaged and functional road surface elements to be identified within a unified framework.

Road surface images were collected through field visits across various rural and urban locations in the Sylhet District of Bangladesh between November 2025 and February 2026. The images were captured using four different smartphone cameras under natural lighting and weather conditions without any artificial setup. This approach was followed to reflect real-world road inspection scenarios and to capture variations in illumination, pavement texture, camera resolution, and viewing angles typically encountered in mobile-based data acquisition systems.

In total, 6,350 road images were initially collected and manually screened to ensure visual clarity and class relevance. Only images in which the primary road surface feature was clearly visible were retained for further processing. Each image was then annotated according to predefined class definitions to maintain labeling consistency throughout the dataset. To address class imbalance and improve variability within individual categories, controlled data augmentation techniques such as brightness adjustment, contrast variation, and blur simulation were applied, resulting in an expanded dataset of 9107 images while preserving the structural characteristics of each class.

The repository includes both the processed (augmented) and original data representations. Specifically, all images used in experiments are provided in a standardized format, resized to a uniform resolution of 1024 × 1024 pixels and stored in .JPG format. The processed dataset consists of augmented images generated through controlled transformations. The original dataset is provided in resized form (1024 × 1024 resolution) rather than in its initial raw capture resolution. Additionally, a metadata CSV file is included, containing structured information such as image path, class label, location, area type, and capturing device, enabling efficient data organization, filtering, and reproducibility of experiments.

The dataset is organized in a structured format suitable for supervised multi-class classification tasks and can be used to benchmark both convolutional and transformer-based models. By incorporating diverse road conditions from rural and urban transportation environments, BDRoad-Sense provides a practical resource for developing and evaluating automated road monitoring systems aimed at improving infrastructure maintenance and transportation safety.

BDRoad-Sense数据集专为支持真实场景下的路面自动化监测与道路安全评估研究而构建。本数据集为多分类图像数据集，涵盖会直接影响行车安全与交通规划的常见路面状态及道路基础设施部件。该数据集包含5个标注类别：严重路面损伤（Major Damage）、轻微路面损伤（Minor Damage）、正常路面（Normal Road）、检查井（Manhole）及减速带（Speed Breaker），可在统一框架下识别受损与完好的路面相关要素。

2025年11月至2026年2月期间，研究团队于孟加拉国锡尔赫特县的多处城乡区域开展实地走访，采集路面图像数据。所有图像均采用四款不同的智能手机摄像头，在自然光照与天气条件下无人工干预拍摄。该采集方案旨在还原真实道路巡检场景，捕捉移动数据采集系统中常见的光照、路面纹理、摄像头分辨率及拍摄视角的各类变化。

初始共采集6350张路面图像，经人工筛选以确保图像清晰度与类别相关性，仅保留核心路面特征清晰可见的图像用于后续处理。随后，所有图像均按照预先定义的类别标准进行标注，以保障全数据集的标注一致性。为解决类别不平衡问题并提升单类别样本的多样性，研究团队采用亮度调整、对比度变换、模糊模拟等可控数据增强技术，最终将数据集扩充至12687张，且保留了各分类的结构特征。

本数据集仓库同时包含预处理（增强后）与原始数据两种版本。具体而言，实验所用图像均采用标准化格式存储：统一调整至1024 × 1024像素分辨率，格式为.JPG。其中，预处理数据集为经可控变换生成的增强图像；原始数据集同样调整至1024 × 1024分辨率，而非采集时的原始分辨率。此外，仓库还附带元数据CSV文件，包含图像路径、类别标签、采集地点、区域类型及拍摄设备等结构化信息，可实现高效的数据组织、筛选及实验复现。

本数据集采用结构化组织形式，适用于监督式多分类任务，可用于卷积神经网络及基于Transformer（Transformer）的模型的性能基准测试。通过涵盖城乡交通环境中的多样路面状态，BDRoad-Sense为开发与评估旨在提升基础设施养护水平与交通安全的自动化路面监测系统提供了实用的研究资源。