Project Description:

How does a robotic medical assistant maintain a steady view of a pulsating artery? How does a self-driving car detect a pedestrian stepping into its path? The answer lies in optical flow – apparent motion of objects in video. But classical algorithms often treat motion as certain, leaving its inherent chaos and uncertainty unaddressed, and critical questions unanswered: How much variation in the robot’s view is introduced by light reflections? Is the car confident enough to brake or swerve in the middle of the road?

We propose a novel optical flow algorithm, which computes a dense optical flow field probabilistically, using iterative extended Kalman filtering to propagate estimates and uncertainties across varying image resolutions. This allows our algorithm to rely on prior motion distributions when image information fails to provide a strong motion signal, and provide confidence estimates that reflect image structure, an essential requirement for safety-critical applications

Project Photo:

Above: First frame of an image pair. Below: its corresponding ground-truth optical flow field.