Qualcomm · Filed May 14, 2025 · Published Jul 30, 2026 · verified — real USPTO data

Qualcomm Patents a Way to Keep Objects Tagged While Building a 3D Map

Building a 3D map of the world is hard enough. Keeping track of which blob of pixels belongs to which chair, person, or car as the camera moves is a whole other problem. Qualcomm's new patent tackles both at once.

Qualcomm Patent: 3D Scene Reconstruction with Object Tracking — figure from US 2026/0220879 A1
Figure from the official USPTO publication.
See all 16 drawings from this filing ↓
Publication number US 2026/0220879 A1
Applicant QUALCOMM Incorporated
Filing date May 14, 2025
Publication date Jul 30, 2026
Inventors Georgi DIKOV, Alin-Eugen DONDERA, Mohsen GHAFOORIAN
CPC classification 345/419
Grant likelihood Medium
Examiner CENTRAL, DOCKET (Art Unit OPAP)
Status Docketed New Case - Ready for Examination (May 24, 2025)
Parent application Claims priority from a provisional application 63750198 (filed 2025-01-27)
Document 20 claims

How Qualcomm's object-aware 3D mapping actually works

Imagine you're wearing AR glasses that need to know exactly where your coffee cup is sitting on the table, not just roughly where the table is. To do that, the device has to build a 3D model of the room and label every object in it, all in real time, as you move your head around.

The tricky part is keeping those labels consistent. When you tilt your head, the camera angle changes, and the software can lose track of which blob of pixels it already identified as "the cup." Qualcomm's patent describes a system that cross-references what the camera sees right now with what it already knows from its existing 3D model, then reassigns labels so nothing gets confused or duplicated.

The result is a 3D scene where every object stays properly identified across frames, even as the viewpoint shifts. That kind of stable, object-level understanding is exactly what AR glasses, robots, and autonomous cameras need to interact meaningfully with the world around them.

How instance maps and depth data update the 3D point cloud

The system works by running two parallel processes on each incoming camera frame and then reconciling them.

Step one: instance segmentation. The processor analyzes the image and produces an "instance map," a pixel-by-pixel label sheet that says "these pixels are object A, those pixels are object B." This is standard computer-vision work, but it's done fresh for every frame, so object IDs can drift or reset.

Step two: projecting the existing 3D model. The system takes the 3D point cloud it has already built (think of it as a dense scatter of labeled dots floating in space) and mathematically "projects" those dots back onto the flat image plane, accounting for the camera's current position and angle (its pose). This creates a second map showing where the already-known objects should appear in this frame.

Step three: reconciliation. The fresh instance map and the projected map are compared. The system reassigns pixel labels from the fresh map to match the stable IDs already in the 3D model, preventing the same object from getting two different names.

Step four: updating the 3D model. A depth map (per-pixel distance data, like what a LiDAR or stereo camera provides) is used to refresh the 3D point cloud with new geometry, and any newly identified object regions are added with their corrected labels.

What this means for AR headsets and autonomous devices

For AR headsets, robotics, and autonomous cameras, stable object identity is a foundational requirement. If a robot's 3D map keeps relabeling the same chair as a new object every few frames, it can't plan a path around it reliably. Qualcomm makes the chips (Snapdragon) that power many of these devices, so a patent like this points directly at on-device processing capabilities it wants to bake into future silicon.

The approach described here is designed to run efficiently on a single device without a cloud connection, which matters for anything that needs low-latency spatial awareness, like a headset tracking your hands or a drone avoiding obstacles. If this ends up in a Snapdragon XR chip, developers building spatial apps would get object-level scene understanding essentially for free.

Editorial take

This is solid, practical computer-vision infrastructure work. It doesn't invent a new sensing modality or model architecture, but it does address a real and annoying consistency problem that plagues real-time 3D reconstruction. For Qualcomm, whose business depends on making AR and robotics workloads run well on its chips, owning this kind of pipeline patent is strategically sensible.

The drawings

16 drawing sheets from US 2026/0220879 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.