SafeWalkBD: A Roadside Object Detection Dataset for Visually Impaired Pedestrians in Bangladesh
Mir Md. Tahmid Kabir, Abdulla Al Mahmud Mugdo, Mohammad Shahidur Rahman · 2024
Roadside objects such as vehicles, persons, and other elements are crucial for developing perception modules in assistive technologies, particularly for visually impaired pedestrians in Bangladesh, where road conditions are complex and unstructured. Existing datasets like MS COCO and KITTI do not adequately capture the unique characteristics of Bangladeshi roads. To address this, we present SafeWalkBD, a high-resolution dataset containing 34,336 images and 16 annotated object classes, including vehicles, traffic signs, poles, potholes, and persons, under diverse environmental conditions. We evaluated the dataset using three object detection models: YOLOv8, YOLOv10, YOLOv11, and Roboflow Object Detection Model 3.0, with YOLOv11 demonstrating the best performance. YOLOv11 achieved a mean average precision (mAP) of 81.2%, with a precision of 82.2% and recall of 75.2%, along with an efficient inference time of 3.04 milliseconds per image. SafeWalkBD aims to enhance mobility and safety for visually impaired pedestrians and to drive further research in object detection and assistive technologies tailored to the unique road conditions of Bangladesh.