Smartphone Based Real Time Cyclist Safety System Using Lightweight YOLOv5n Object Detection
Year: 2025
Type: article
Abstract: Cyclist fatalities are increasing globally—1,155 deaths were recorded in the US alone in 2023. Many of these incidents involve single-bicycle crashes caused by road hazards such as potholes, debris, or animals. This paper presents a smartphone-based real-time hazard detection system designed to enhance cyclist safety by providing timely warnings using a lightweight AI model. The system leverages YOLOv5n, an ultra-efficient object detection model optimized with TensorFlow Lite and OpenCV. YOLOv5n was selected over other lightweight models like MobileNet SSD or YOLOv6-lite for three reasons: (1) it offers a superior balance of inference speed and detection accuracy on mid-range smartphones; (2) its modular PyTorch implementation and quantization pipelines are deployment-friendly; and (3) it consistently outperformed MobileNet SSD in detecting small hazards while using less memory than YOLOv6-lite. The system detects both static (e.g., potholes, debris) and dynamic (e.g., vehicles, pedestrians) hazards from live video streams and issues alerts. Trained on a curated dataset of 2,000+ images, it achieved a mean Average Precision ([email protected]) of 0.86 and operated at 12.8–13.4 FPS on mid-range smartphones. Field tests showed an alert lead time of 1.2 ± 0.2 seconds, enhancing rider response and braking time. This can reduce impact risk by over 35% in urban cycling conditions [1]. Unlike server-based or embedded alternatives, this privacy-preserving solution runs entirely on-device. Aligning with Vision Zero and sustainable mobility goals [2], it can integrate into municipal hazard systems or cyclist insurance incentives, fostering both safety and behavioral accountability.
Author: Gautam Rulyan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4412595869