Sangmin Hong won the 1st place in the Zero-Shot Accident Anticipation Challenge at the CVPR 2026 AUTOPILOT Workshop.
The proposed method presents a training-free framework for anticipating accident risk from driving videos by leveraging pretrained video understanding models, semantic scene descriptions, and temporal risk aggregation.
Without task-specific accident annotations or model training, the approach identifies emerging hazardous situations from multimodal driving inputs and demonstrates strong generalization across diverse real-world traffic scenarios.
Workshop homepage and Leaderboard can be found at the links below:
[Homepage] [Leaderboard]
Congratulation!
Recent Comments