Google Patents a Video Compression Method That Predicts Object Motion Between Frames
Every compressed video hides a mountain of math, and a big chunk of that math is just describing where things moved. Google's new patent tries to make that math cheaper by guessing the movement before it even looks at the current frame.
How Google's motion-path prediction trims video file sizes
You're watching a soccer match on your phone and the camera pans across the field. Every second, your device is decoding dozens of frames, and for each one it has to figure out how every part of the image shifted from the last frame. That movement data takes up space in the video file, and squeezing it down is one of the central challenges of video compression.
Google's patent describes a method where the encoder and decoder both trace the movement paths of objects across several already-processed frames, building a kind of motion map. When the next frame arrives, the system uses that map to predict where everything has moved, so it only needs to store the small correction if the prediction was off.
The result is that the video file carries less redundant information about motion, which can mean smaller file sizes, faster streaming, or better picture quality at the same bandwidth. You'd feel the difference as fewer artifacts during fast action, or smoother playback on a slow connection.
… obtain a motion vector prediction for a current block from the current frame in accordance with the trajectory mapping data …
Translation: It estimates where an object will move in the current frame using past movement data.
How the encoder maps trajectories across reference frames
The patent describes a system for both encoding (compressing) and decoding (playing back) video that uses motion vector prediction with derived motion trajectory.
A motion vector is simply a pair of numbers that says "this block of pixels moved X pixels left and Y pixels down since the last frame." Storing those vectors for every block in every frame is expensive. The standard shortcut is to predict the vector from neighboring blocks and then only store the error (called a differential motion vector). Google's method improves the prediction step.
Here is how the pipeline works:
- The encoder looks at previously reconstructed reference frames (frames already processed) and extracts their motion field data, essentially a map of how things were moving.
- It uses that map to compute trajectory mapping data, tracing the path a moving object has followed across multiple past frames.
- It then projects that trajectory forward to predict where the object will be in the current frame, producing a motion vector prediction.
- The actual motion vector for the current block is compared to the prediction; only the small difference (the differential) is stored in the compressed bitstream.
The decoder runs the same trajectory-reconstruction logic independently, regenerates the same prediction, adds the stored differential back, and arrives at the correct motion vector to reconstruct the frame. Because both sides can derive the trajectory from already-shared data, no extra trajectory information needs to travel in the stream.
… obtaining a motion vector for the current block by adding the motion vector prediction and the differential motion vector …
Translation: It calculates the final motion by combining the predicted movement with any correction data.
What this means for streaming quality and storage costs
For anyone who watches a lot of video, the downstream effect of better motion prediction is either lower data usage for the same quality or higher quality at the same data rate. Fast-moving content like sports, gaming streams, and action films tends to stress motion-prediction systems the most, so those are the cases where an improvement here would be most noticeable to you.
Google's long investment in open video-codec work (it is a lead contributor to the AV1 and AV2 codec projects) suggests this kind of research feeds directly into codecs that ship in Chrome, YouTube, and Android. A more efficient motion model baked into a future codec revision could improve video quality for billions of streams without any change visible to users.
That makes this Google's 688th filing in our Google coverage since May, a corpus that spans work like skin-tone heart-rate sensing and Waymo collision prediction.
If you have ever watched a video call hold its quality for an extra second when your connection dips, or noticed a streaming video recover its sharpness a little faster than expected, this is the sort of work behind that moment. Google's patent describes a smarter way for video software to predict where moving objects will be in the next frame, by studying how they moved across several recent frames rather than just guessing from the last one.
Because the prediction is more accurate, the video file wastes less space correcting wrong guesses and keeps more room for actual picture detail. That tradeoff benefits everyone watching video on a strained connection or a limited data plan.
This will never appear in a feature list. It lands years from now as a small, invisible improvement baked into a standard that makes your next video call a little cleaner or your next stream a little lighter on your phone bill.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Google patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
14 drawing sheets from US 2026/0270441 A1 · click any drawing to enlarge
Show the other 6 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in