Get the newsletter
Microsoft · AI / ML · Filed Apr 6, 2026 · USPTO published Aug 13, 2026 · Aug 13 release

Microsoft Patents a Way to Spot People in Video Before Even Opening the File

Microsoft has filed a patent for a system that can detect whether a person appears in a video stream without ever fully decoding the footage, a trick that could dramatically cut the computing cost of video analysis at scale.

Video frames captured by a security camera mapped alongside a packet size graph for compressed data analysis. Drawing from patent filing US 2026/0238800 A1.
Video frames captured by a security camera mapped alongside a packet size graph for compressed data analysis.All 8 drawings ↓
Publication number
US 2026/0238800 A1(US20260238800A1)
Applicant
Microsoft Technology Licensing, LLC
Filed · Published
Apr 6, 2026 · Aug 13, 2026
Inventors
Yichen HAO, Lihang LI, Anthony C. ROMANO, Naiteek SANGANI, Ryan S. MENEZES
Status when we published
Waiting for an examiner
Document
23 claims
US classification
375/240.02
Parent application
Division of 17/968,694 (filed Oct 18, 2022)(17968694)

What Microsoft's frame-skipping detection actually does

You're watching a recorded security feed, and somewhere on a server, software is scanning thousands of hours of footage to find the moments a person walks into frame. Normally that means decoding every single video frame, which is computationally expensive and slow.

Microsoft's patent describes a shortcut: scan the encoded video data directly, before it's ever unpacked into actual images. Encoded video already contains compressed clues about what's in each frame, and the system learns to read those clues well enough to flag frames that likely contain a person, skipping the heavy decoding work entirely.

The result is that a server could process video using a fraction of the usual resources, or check far more footage in the same amount of time. That matters any time you need to find a person in a large library of video without having a human watch all of it.

From the filing · THE ABSTRACT
The description relates to resource aware object detection for encoded video streams that can identify frames of the video stream that include an object of interest, such as a human, without decoding the frames.

Translation: Microsoft patented a method to find people in videos without actually playing or opening the files first.

How the system reads encoded data without decoding frames

Standard video analysis works like this: take a compressed video file, decode each frame into a full image, then run an object-detection model on that image to check whether a person (or any object) appears. It works well but burns through CPU and GPU resources fast, especially at scale.

Microsoft's patented approach cuts out the decoding step for the initial scan. Encoded video formats like H.264 or H.265 store data as a mix of compressed signals, including motion vectors (which describe how pixels move between frames) and residuals (the leftover differences after motion is accounted for). These compressed representations still carry meaningful structural information about scene content.

The system trains a detection model directly on this encoded-domain data rather than on decoded images. When a new video stream arrives:

  • The system reads the encoded bitstream without decompressing it into frames.
  • The model inspects the compressed signals for patterns associated with human figures.
  • Frames flagged as likely containing a person can then be selectively decoded for full analysis.

The key tradeoff is accuracy versus speed: the encoded-domain model is probably less precise than a full-image detector, but it works as a fast filter that eliminates obviously empty frames before the expensive step begins.

What this means for Teams, Azure, and real-time video AI

For any service that processes video at scale, decoding is one of the biggest infrastructure costs. Microsoft runs Teams meetings, Azure Video Indexer, and a range of enterprise surveillance and media tools, all of which involve scanning video for people or objects. A system that can triage frames before decoding them could meaningfully reduce server load, which translates to lower costs and faster turnaround on video search tasks.

From a ship-path perspective, this is software-only work: no new hardware is required, and the core idea could slot into existing video-processing pipelines relatively quickly, assuming the encoded-domain model reaches acceptable accuracy thresholds. The latest Big Tech patents in the video-AI space show a clear push toward inference that happens earlier in the media pipeline, and this filing fits squarely in that direction.

Source: US 2026/0238800 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

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 Microsoft 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 · 8 sheets

8 drawing sheets from US 2026/0238800 A1 · click any drawing to enlarge

Video frames and packet graph, drawing sheet 1 from US 2026/0238800 A1
Sheet 1 Video frames and packet graph
Video frames with detection thresholds, drawing sheet 2 from US 2026/0238800 A1
Sheet 2 Video frames with detection thresholds
Video processing pipeline architecture, drawing sheet 3 from US 2026/0238800 A1
Sheet 3 Video processing pipeline architecture
Flowchart for resource expenditure, drawing sheet 4 from US 2026/0238800 A1
Sheet 4 Flowchart for resource expenditure
Flowchart for group processing, drawing sheet 5 from US 2026/0238800 A1
Sheet 5 Flowchart for group processing
Flowchart for metadata comparison, drawing sheet 6 from US 2026/0238800 A1
Sheet 6 Flowchart for metadata comparison
Flowchart for packet size analysis, drawing sheet 7 from US 2026/0238800 A1
Sheet 7 Flowchart for packet size analysis
Network of cameras and servers, drawing sheet 8 from US 2026/0238800 A1
Sheet 8 Network of cameras and servers
Patent filing page