Enhanced Cyclist Safety Through Vision-Language Models: A Real-Time Advisory Framework

Cycling safety is a priority with rising road accidents. This study establishes an annotation scheme for cycling videos and integrates a vision-language model (VLMs) with real-time safety recommendations. Using 11 cycling videos depicting diverse Queensland environments, the study analyzes thousands of video frames at a rate of 1 frame per second, applying Zero-Shot in-context learning with a multimodal language model. The system generates structured safety insights based on contextual input from Queensland Road Rules and assigns a safety score out of 10 per image. Initial evaluations highlight the system's effectiveness in identifying rule violations and generating real-time advice for compliance with designated bike lanes and defensive riding strategies. However, occasional misclassification of traffic elements, such as brake lights being interpreted as signals, highlights the need for improved resolution and refinement of the model's contextual understanding. The AI-based approach keeps cyclists attentive, encourages defensive cycling, and reduces global cycling accident rates.