Rear Collision Warning Detection System for Bicycle Using YOLOv8n and Adaptive Kalman Filter: A Vision-Based Approach with Adaptive Speed Estimation

Cyclists in urban environments face significant risks from rear-end collisions due to limited awareness of approaching vehicles. Traditional safety measures, such as rear-view mirrors and wearable sensors, often fail to provide timely and precise warnings, necessitating advanced technological interventions. Recent advances in computer vision and embedded systems have enabled vision-based collision detection systems, but these often rely on expensive sensors or lack realtime adaptability. To address these challenges, this paper introduces a vision-based rear collision warning system for bicycles, utilizing the lightweight YOLOv8n model for real-time object detection and DeepSORT for robust multi-object tracking. The system estimates vehicle distance via monocular vision and predicts trajectories using a Kalman filter to compute the Time-To-Collision (TTC), issuing warnings based on TTC and proximity thresholds. Without the need for real-time GPS, an adaptive Kalman filter is proposed to estimate bicycle speed from video overlays, validated separately with simulated data. Evaluated using repurposed rear car camera footage, the system achieves high detection accuracy and tracking consistency. This work advances cyclist safety by integrating state-of-the-art computer vision into a cost-effective, deployable solution, overcoming limitations of prior sensor-heavy systems.