RULES · CONTEXT · UNDERSTANDING
ROAD.
The same action.
A different meaning.
Integrating driver video, driving context, and behavioral rules to identify previously unseen driver anomalies—and explain why a behavior is abnormal.
Research contributions
- Designed rule-grounded reasoning that translates domain knowledge into structured guidance for multimodal large language models.
- Integrated context-aware visual reasoning to distinguish similar behaviors across driving situations and account for context-dependent safety rules.
- Built an evaluation pipeline across driver-monitoring datasets with free-form anomaly descriptions, human assessments, and MLLM baselines.
When safely parked, looking at a phone does not conflict with the active task of driving.
Illustrative scenario · manually defined responses, not model predictions.
