CIRS: A Multi-Agent Machine Learning Framework for Real-Time Accident Detection and Emergency Response.

Area of Science:

• Computer Science
• Artificial Intelligence
• Traffic Safety Engineering

Background:

• Road traffic accidents are a major global cause of death, with delays in detection and emergency response exacerbating outcomes.
• Existing machine learning accident detection systems struggle with complex modern traffic environments and lack widespread deployment.

Purpose of the Study:

• To introduce Collaborative Intelligence for Road Safety (CIRS), a novel multi-agent framework for real-time accident detection, semantic scene understanding, and coordinated emergency response.
• To enhance situational awareness and response efficiency in road traffic scenarios through a collaborative multi-agent approach.

Main Methods:

• Developed CIRS, a decentralized, multi-agent machine learning framework with specialized agents for perception, classification, description, localization, and decision-making.
• Integrated YOLOv11 for high-accuracy accident detection and VideoLLaMA3 for rich semantic scene description.
• Evaluated performance on a custom dataset of 10,000 frames (5200 accident, 4800 non-accident).

Main Results:

• YOLOv11 achieved 86.5% top-1 and 100% top-5 accuracy for reliable real-time accident detection.
• VideoLLaMA3 demonstrated superior factual accuracy and reduced hallucinations compared to other vision-language models, with BLEU (0.0755), METEOR (0.2258), and ROUGE-L (0.3625) scores.
• The multi-agent architecture proved scalable, reduced latency, and minimized false positives.

Conclusions:

• CIRS effectively bridges low-level visual analysis with high-level situational awareness for improved road safety.
• The framework enables timely emergency service dispatch through enhanced detection and scene understanding.
• CIRS offers a scalable and efficient solution for real-time accident detection in complex traffic environments.