CDS&E: Collaborative Intelligence Adaptation of Multimodal Data Utility Theory and Foundation Models for Traffic Safety Interventions

This Computational and Data-Enabled Science and Engineering (CDS&E) project addresses the urgent need for research communities and transportation agencies to better understand traffic crash causation and translate safety data into effective safety improvements. It aims to transform traffic safety analysis and engineering decision-making by harnessing the collaborative intelligence of multimodal data, generative AI, and data utility theory. The core of the project is the SafeTraffic Decision Copilot, a decision engine designed to predict event-level crash risk, explain contributing factors, and recommend targeted safety interventions and future data collection strategies. The project advances methods for crash event prediction using multimodal foundation models, crash knowledge graphs, uncertainty quantification, and causal “what-if” analysis informed by key features of roadway infrastructure, driver behavior, vehicle attributes, weather, and crash narratives. In collaboration with transportation stakeholders, the research team will support the practical translation of these methods into agency-facing safety analysis and decision-support workflows. The novelty of the SafeTraffic Decision Copilot lies in building collaborative intelligence across data, models, and decisions for trustworthy traffic safety analysis and intervention. The project is organized around three main thrusts: 1) Knowledge adaptation: to advance foundation model task adaptation for crash analysis, the project develops crash knowledge graphs and knowledge-informed training methods that integrate roadway, driver, vehicle, weather, and crash narrative information into generative artificial intelligence models. 2) Data adaptation: to improve the use of fragmented and heterogeneous safety data, the project advances multimodal data utility theory to quantify the value of data samples, sources, and combinations, enabling efficient data selection, robust data mixing, and scalable model training. 3) Decision adaptation: to translate predictions into safety actions, the project develops transparent crash-risk attribution, uncertainty-aware prediction, and alternative-scenario analysis to identify contributing factors and recommend targeted engineering interventions. If successful, the project will provide a critical knowledge base for shifting traffic safety practice from reactive crash analysis to proactive crash prevention. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.