Abstract
Road traffic accidents remain a leading cause of fatalities worldwide, and the consequences are considerably worsened by delayed detection and emergency response. Although several machine learning-based approaches have been proposed, accident detection systems are not widely deployed, and most existing solutions fail to handle the growing complexity of modern traffic environments. This study introduces Collaborative Intelligence for Road Safety (CIRS), a novel, multi-agent, machine-learning-based framework designed for real-time accident detection, semantic scene understanding, and coordinated emergency response. Each agent in CIRS is designed for a distinct role perception, classification, description, localization, and decision-making, working collaboratively to enhance situational awareness and response efficiency. These agents integrate advanced models: YOLOv11 for high-accuracy accident detection and VideoLLaMA3 for contextual-rich scene description. CIRS bridges the gap between low-level visual analysis and high-level situational awareness. Extensive evaluation on a custom dataset comprising (5200 accident, 4800 nonaccident) frames demonstrates the effectiveness of the proposed approach. YOLOv11 achieves a top-1 accuracy of 86.5% and a perfect top-5 accuracy of 100%, ensuring reliable real-time detection. VideoLLaMA3 outperforms other vision-language models with superior factual accuracy and fewer hallucinations, generating strong results in the metrics of BLEU (0.0755), METEOR (0.2258), and ROUGE-L (0.3625). The decentralized multi-agent architecture of CIRS enables scalability, reduced latency, and the timely dispatch of emergency services while minimizing false positives.