Vision-Based Real-time Hazard Identification and Alerting for Rider Assistance

Electric bicycles (e-bikes) have gained significant popularity due to their eco-friendliness, affordability, and accessibility. Moreover, e-bikes serve as an inclusive transportation solution for individuals with limited physical stamina or mobility constraints, enabling longer and more frequent travel than traditional bicycles.This paper introduces an Advanced Rider Assistance System (ARAS) designed to enhance cyclist safety on smart e-bikes. The system is implemented on a prototype smart e-bike platform developed at the German University in Cairo (GUC), equipped with a resource-constrained Raspberry Pi 4. It leverages real-time embedded vision to detect road hazards.The proposed ARAS builds upon previous work in pothole detection, speed bump recognition (both marked and unmarked), cyclist alerting mechanisms, and traffic sign interpretation. A custom dataset was collected to train and evaluate several state-of-the-art object detection architectures, including YOLOv11, YOLOv12, RT-DETRv2, and D-FINE. Our findings show that the transformer-based D-FINE-N model provides the optimal performance-speed trade-off, demonstrating a superior balance of precision (88.7%) and recall (87.5%) and a leading F1-score of 88.1% and mAP@50-95 of 49.3%. Moreover, after optimization and conversion to the ONNX format, D-FINE-N achieved the highest inference speed, reaching 2.75 FPS on the target hardware. This research demonstrates the viability of deploying modern deep learning models for ARAS applications on low-power edge devices and highlights the effectiveness of D-FINE-N for real-time safety systems.