Motion and Tracking

videos add dimension of time

• optical flow

• lucas-kanade - pyramids for large motion

• horn-schunk

• segmentation

• feature tracking

• slam

Challenges

• All the points on an edge look the same. It is hard to estimate where each point will move to.
• Low-texture regions have small eigenvalues. The matrix is harder to invert and get accurate estimates of optical flow
• high texture regions have large lambda1 and lambda2

Fixes