Adaptive object detection in dynamic road environments using YOLOv11 with continuous learning for Real-Time obstacle detection

Abstract

Abstract Real-time object detection in dynamic road environments is crucial for enhancing road safety, infrastructure monitoring, and Intelligent Transportation Systems (ITS). This study presents an adaptive object detection model based on YOLOv11 with continuous learning, designed to detect road obstacles in real-time without frequent retraining autonomously. Unlike existing YOLO-based models such as YOLOv10 and LD-YOLOv10, the proposed model integrates a continuous learning mechanism, enabling it to adapt dynamically to evolving road conditions, including potholes, vehicles, signposts, and other obstacles. The model was trained and evaluated on a dataset collected from diverse urban road environments in Uganda, achieving a mean average precision (mAP) of 0.856, precision of 0.873, and recall of 0.849. It also demonstrates high-speed performance at 78.7 Frames Per Second (FPS), making it suitable for real-time deployment on resource-constrained edge devices. While this study primarily focuses on model development and optimization for road obstacle detection, future research will explore its potential application in assistive navigation systems for visually impaired individuals. Comparative analysis against state-of-the-art models, including YOLOv10, LD-YOLOv10, and Tiny-DSOD, demonstrates the competitive performance of the proposed model, particularly in detecting small and occluded objects.