Multimodal System Identifies Collisions, Near-misses in Traffic Videos

C2SMART researchers led by Professor Kaan Ozbay developed SeeUnsafe, a system using multimodal large language models to automatically identify collisions and near misses in traffic videos. Tested on the Toyota Woven traffic safety data set, the model correctly classified videos showing collisions, near misses, or normal traffic 76% of the time and identified specific road users involved in incidents up to 87% of the time. The system generates reports detailing factors leading to traffic events, enabling transportation officials to implement corrective solutions like improved signage or redesigned road layouts.