optical flow
/ OP-tih-kul floh /
Optical flow measures motion between two consecutive video frames by asking, for every pixel: where did it move to? The output is a dense field of tiny arrows, one per pixel, showing the direction and speed of apparent movement. Imagine flipping between two photos taken a fraction of a second apart and tracing how each speck of the scene slid across the frame — that map of slides is optical flow. It is how a machine perceives that things are moving and roughly how fast.
Classically, the method rests on a simple assumption: a patch of the world keeps the same brightness from one frame to the next, so you can find where it went by matching brightness patterns. From that you can stabilize shaky video, compress video (store the motion instead of re-storing every pixel), interpolate smooth slow-motion between frames, track objects, and feed motion cues into action-recognition and robotics. Newer systems learn flow with neural networks and handle harder cases better.
The honest limits are baked into the name: it is optical flow, the apparent motion of brightness in the image, not the true motion of objects in the world. A spinning barber pole seems to move upward; a uniform white wall sliding past gives no flow at all because there is nothing to track; shiny reflections and shadows move without any object moving. The classic brightness assumption also breaks under lighting changes, large fast motions, and occlusion. So flow is a powerful motion cue, but it is an inference about appearance, not a direct reading of physics.
Two frames of a car driving left to right: optical flow draws rightward arrows densely across the car body, longer arrows where it moves fastest, and near-zero arrows on the still background. The smooth, evenly painted car door, having no texture to track, may get unreliable or missing arrows in its middle.
Arrows per pixel show apparent motion — but texture-less surfaces give no signal to track.
Optical flow is the motion of brightness in the image, not of objects in the world — reflections, shadows, and lighting changes all create flow with nothing actually moving, and a featureless surface that is moving creates none. It is a cue to be interpreted, not a measurement to be trusted blindly.