Beyond just keeping hands on the wheel: Towards visual interpretation of driver hand motion patterns

Observing hand activity in the car provides a rich set of patterns relating to vehicle maneuvering, secondary tasks, driver distraction, and driver intent inference. This work strives to develop a vision-based framework for analyzing such patterns in real-time. First, hands are detected and tracked from a monocular camera. This provides position information of the left and right hands with no intrusion over long, naturalistic drives. Second, the motion trajectories are studied in settings of activity recognition, prediction, and higher-level semantic categorization.