Project R-15839

Leveraging Generative AI for Integrated Traffic Accident Scene Analysis and Reconstruction (Research)

Post-crash investigation helps in determining the accident causes to prevent future crashes. Various factors contribute to accidents, including human error, road condition, and environmental factors. Accurate analysis requires extensive data collection from vehicle movements, road conditions, eyewitness accounts, and police reports. Accident reconstruction helps recreate incidents using methods like numerical modeling, simulations, and crash data retrieval. Traditional accident analysis methods face limitations such as manual processing, subjectivity, and privacy concerns. This study utilizes Generative AI and Multimodal Large Language Models (MLLMs) for accident reconstruction and analysis. MLLMs combine textual and non-textual data, integrating advanced language models like GPT with other models such as vision models for more accurate predictions. The foundation of MLLMs build on transformer architecture that uses an attention mechanism. They undergo pre-training on large datasets and fine-tuning for domain-specific tasks, improving efficiency and reducing costs. In this study, data will be collected from relevant organizations such as police and insurance companies, followed by data fusing and fine-tuning of open-source, pre-trained models. A tool will be developed utilizing the fine-tuned model which will be evaluated using statistical and machine learning metrics. This tool will facilitate the authorities and insurance companies for accident reconstruction and analysis.

19 June 2025 - 30 June 2028