UWB Plus Overhead Cameras: Sub-30 cm Forklift Tracking, Fewer Anchors

Why does pure UWB miss 30 cm on a real forklift?

An 802.15.4z UWB radio ranges to about 10 cm in a certification lab. A warehouse is not that lab. Steel racking blocks the direct path, the receiver locks onto a reflection, and a non-line-of-sight range reads long by anywhere from half a meter to several meters. TDoA positioning needs four anchors with a clean view of the tag at the same instant, and in a racked aisle the only clean view is a narrow canyon straight up.

Two consequences follow:

• Anchor density explodes. An open-ceiling grid with 25 to 30 m spacing collapses into per-aisle anchor placement, because an anchor two aisles over is behind three rows of steel and shrink-wrapped pallets.
• Update rate fights physics. At a typical 10 to 12 km/h site limit, a truck covers about 3 m every second. At a 2 Hz position rate it moves more than 1.5 m between fixes, which alone blows a 30 cm budget on a moving target.

You can fight both with more anchors and faster ranging, and many deployments do, at a PoE drop, a conduit run, and scissor-lift hours per anchor. The spec sheet assumed an empty room.

What does an overhead camera add that a fifth anchor cannot?

A fifth anchor gives you one more range with the same multipath problems as the first four. A camera gives you different physics. A 4K sensor mounted at 10 m, covering roughly 18 by 10 m of floor, resolves under 1 cm per pixel. After lens calibration and a floor-plane homography, a detected forklift converts to floor coordinates within a few centimeters, limited by detection jitter rather than optics.

The camera also reports what no radio does:

• Heading, from vehicle orientation, so you know which way the forks point before the truck moves.
• Load state, forks loaded or empty, which turns a position log into a workflow record.
• Untagged objects, pedestrians and dropped pallets that a tag-based system is blind to.

The weaknesses are just as specific. Vision cannot tell two identical reach trucks apart, and it loses targets behind racking and in glare off stretch wrap. UWB is the mirror image: the tag ID is never ambiguous, and ranging keeps working at blockage levels that end a camera track. Each one fails where the other is strong.

How does the fusion actually work?

Run one estimator per truck, an extended Kalman filter or a small factor graph, with a state of position, velocity, and heading. Feed it two measurement types. Camera detections enter as floor-plane position plus heading. UWB enters as raw ranges to known anchor positions, not as a pre-computed fix. That distinction matters: a TDoA fix needs four anchors at once, but a single raw range is useful on its own, so even two reachable anchors keep tightening the estimate.

Association is the part teams underestimate. When a camera track and a tag trajectory overlap for a second or two, bind the track to the tag ID. From then on, identity rides through camera handoffs and camera precision rides through UWB dead spots.

Two implementation rules:

• Sync clocks to within 10 ms. A forklift at 3 m/s moves 3 cm in 10 ms, a tenth of the whole budget.
• Gate NLOS ranges. If a range disagrees with the camera-backed prediction by more than your gate, it is a reflection. Drop it.

Fuse measurements, not finished positions.

How much hardware does fusion actually remove?

Pure UWB TDoA in a racked facility is a density problem. Once steel forces per-aisle placement, plan on one anchor per 1,000 to 2,500 square feet, which is 40 to 100 anchors per 100,000 square feet. The radio might cost $200; the installed point rarely stops there. A PoE drop, conduit, a scissor lift, and night-shift electrician hours commonly push each point past $1,000, and that line item scales with every anchor.

Fusion redraws the coverage map by zone:

• Docks, staging, cross-aisles: clear ceiling sightlines, so cameras carry the accuracy here. One overhead 4K camera covers on the order of 180 square meters of floor and can often reuse existing security-camera mounts and PoE runs.
• Racked aisles: keep UWB, but because raw ranges feed the filter, two reachable anchors do work TDoA needed four for.

Facilities designed this way commonly land near half the anchor count of a pure UWB plan at the same accuracy target. Cable and lift time cost more than radios.

What accuracy spec should you write into the contract?

Write the spec against the worst case, because vendors quote the best one.

• Accuracy: CEP95 under 30 cm, measured in your tightest racked aisle during a live shift, not a facility-wide average taken on a quiet Sunday. Full racking and trailers at the docks change the RF environment.
• Heading: within 5 degrees, if you want direction-of-travel alerts rather than dots on a map.
• Latency: under 100 ms end to end if positions feed pedestrian warnings; one second is fine if they only feed heatmaps and slotting analytics.
• Update rate: 5 Hz or better on moving vehicles. At 3 m/s, a 1 Hz system means 3 m of blind travel between fixes.

For acceptance, lay out taped waypoints with a laser distance meter, then drive the real route at working speed with the mast raised and loaded, because a raised steel mast shadows a cab-mounted tag. Test the worst aisle, loaded, at speed.

Should you build the fusion stack or license it?

The Kalman filter is a textbook chapter. The product is everything around it: binding camera tracks to tag IDs when three trucks cross paths, rejecting NLOS ranges without rejecting real turns, recalibrating after a mast clips a camera mount, and keeping a 60-truck fleet correctly associated across shift changes. Teams that build this from scratch typically spend a year getting from a clean-room demo to numbers that hold up in a live facility.

There is also a freedom-to-operate question. RF ranging fused with optical tracking is heavily patented ground, and discovering that after launch is the expensive way to learn it. Position Imaging licenses hundreds of granted patents across real-time positioning, RF ranging, computer vision, and machine learning, including US 11,774,249, US 12,079,006, US 12,066,561, and US 12,000,947, a portfolio cited by Apple, Bosch, and other major firms in their own filings. A license covers both sides at once: methods you do not have to re-derive, and standing to practice them. License the solved parts and ship.