Motion detection
Also known as: movement detection, activity detection
Measuring how much is moving in a stretch of footage, the cheapest and most reliable first filter for finding anything interesting.
Updated
What it means
Motion detection answers a blunt question: is anything happening here. It compares frames and reports how much changed, which separates a rider dropping into a feature from four minutes of a camera pointed at a chairlift.
It understands nothing about what is moving, and that is exactly why it is dependable. There is no training data to be unrepresentative and no model to be confidently wrong, only arithmetic.
In a highlight pipeline it is usually the first pass, because it can run over hours of footage without any of the input limits that constrain a Multimodal model working on the same material.
How it works in practice
The measurement comes from comparing successive frames, or from data the codec already stores, which Motion vector analysis reads straight out of the file. The second route is far cheaper because the encoder did the work during recording.
Output is a curve over time, and its peaks are candidates. Those candidates then go to something that can tell a jump from a shaky walk, which is where Action recognition takes over.
Smoothing the curve before reading peaks avoids a thousand false candidates from ordinary camera shake, and takes no measurable time to apply. A minimum gap between peaks does the same job for a single long burst of activity.
What to watch out for
Camera movement reads as subject movement. A chest mounted camera on a rough trail produces constant high motion with nothing happening, so raw motion alone ranks the bumpiest footage highest.
The opposite error loses the quiet shot. A deliberate static landscape measures as nothing and is dropped, along with the breathing space it would have given the edit between two impacts.
Light changes also masquerade as motion. A cloud crossing the sun shifts every pixel at once, which looks identical to movement unless the measurement separates whole frame changes from local ones. A sudden Exposure shift when a camera leaves a tunnel produces the same false spike.
How RawClip handles it
Motion is measured across an entire upload first, because it is cheap enough to run on everything, and only the stretches that register go on to the slower analysis that can judge what the movement actually was.