SynSHRP2: A Synthetic Multimodal Benchmark for Driving Safety Critical Events Derived from Real-World Driving Data

Project Description SynSHRP2 is a large-scale synthetic multimodal dataset derived from the SHRP 2 Naturalistic Driving Study (NDS), focusing on safety-critical driving events such as crashes and near-crashes. The dataset includes over 6,500 events (1,340 crashes and 5,191 near-crashes) with detailed annotations on event types, conflict types, and incident types. It provides synchronized multimodal data including time-series vehicle kinematics, narrative descriptions, and de-identified synthetic keyframe images generated using advanced AI techniques (Stable Diffusion, ControlNet, IP-Adapter) to preserve privacy while maintaining critical scene information. The dataset supports research in crash prediction, scene understanding, and autonomous driving safety. Data Types Time-series sensor data (vehicle kinematics, driver inputs) Synthetic RGB keyframe images (1920×1080 resolution) at critical event timestamps Narrative text annotations describing scene context and environmental conditions Tabular Records (categorical labels for event, conflict, and incident types) Data Format Numerical time-series data in structured JSON files Images in PNG formats Text annotations in structured JSON files Accompanying data dictionary detailing variable definitions and coding schemes A comprehensive data dictionary detailing variable definitions, data types, and coding schemes is provided as an accompanying document.