Advancing autonomous navigation: YOLO-based road obstacle detection and segmentation for Bangladeshi environments

Abstract

The advancement of autonomous vehicles requires a fast and effective object detection and segmentation system handling a wide range of road environments. The goal of this research is to improve autonomous navigation by applying fast and popular YOLO-based models for road obstacle detection and segmentation in South Asian countries, especially Bangladesh. Our team has compiled an extensive collection of videos taken on Bangladeshi streets using a smartphone camera that shows a variety of road conditions such as potholes, speed bumps, barricades, and normal roads taking into account rainy, sunny, day and night environments. By using Roboflow annotation and sampling tools, these videos were sampled into images and annotated with both bounding boxes and bounding masks. Using our custom annotated dataset, we trained and refined YOLO-based object detection and segmentation models such as YOLOv5, YOLOv7, and YOLOv8. The YOLOv5x model trained on our custom dataset shows better results with the highest mAP50 and mAP50-95 scores of 0.876 and 0.647 respectively. However, the YOLOv7x model trained on our custom dataset gives the lowest performance outcome with mAP50 and mAP50-95 scores of 0.583 and 0.331. Also, while comparing the models trained with custom dataset with models trained with a benchmark dataset, our dataset shows major improvements in results. We deployed the models on our local computer to allow real-time object detection with a camera, upon which our prototype car can take decisions using a microprocessor. This implementation reflects the feasibility of effective object detection and segmentation with limited resources. This research intends to optimize autonomous vehicle navigation in Bangladeshi road environments quickly and effectively. The outcomes of our experiments suggest that our approach offers a viable way to improve the security and efficiency of autonomous navigation in these kinds of settings. By addressing the unique challenges with road infrastructure in developing nations, our research advances the area of autonomous driving and creates opportunities for more customized and adaptive navigation systems.