Skip to content

Optical FLow

Programming / vision

Optical flow is the estimation of the apparent motion of objects or pixels between consecutive frames in a video, based on how the brightness pattern changes.

  • Input → two frames (images)
  • Output → displacement vectors (u,v) showing where each pixel (or feature) moved

Types

  • Gradient-based (Lucas–Kanade, Horn–Schunck)
  • Feature-based (track sparse keypoints)
  • Dense methods (Farnebäck, TV-L1, Brox)
  • Deep learning–based (FlowNet, RAFT).

Dense vs sparse

Dense

Motion is computed fro every pixel in the image

Sparse

Motion is computed only for selected points (usually features like corners or blobs).

OpenCV Lucas–Kanade OpenCV Lucas–Kanade cuda implementation