Roadmarker

// 01. The problem

A backup camera shows you a picture. Driving something large needs a volume: what is solid, where it ends, and how much room is left. The wider the vehicle and the worse the sightlines, the more of that volume you are guessing at.

Roadmarker takes six cameras mounted around the vehicle and reconstructs the space as occupied and free, then tracks what moves through it and holds identity on each thing across frames. A road marker is what makes a route followable when the road itself will not tell you where it goes.

// 02. What it looks like scene-0103 · frame 057

The Roadmarker display showing a reconstructed intersection as a
blue point cloud, with 25 tracked vehicles and pedestrians
labelled by class, confidence and range, the vehicle's own
footprint ringed in green, and the forward camera inset at
centre. — 25 tracked objects, 835,613 points, 13 fps. Each label carries what the thing is, how sure the system is, and how far away it sits. The green ring is your vehicle.

// 03. Where it fits three setups

Vans & RVs

Park a long wheelbase in a real space. See the bollard, the low branch and the drop at the back of the site without leaning out of a window.

Overland

Pick a line over rock and ruts with the ground under the sills drawn in, including the part the hood hides.

Fleet

Cut low-speed damage on box trucks and delivery vans, and keep a record of what the vehicle could actually see.

// 04. The kit as planned

Included

Compute unit

A sealed box that runs the perception stack on the vehicle. No cloud in the loop, because a system you rely on while reversing cannot wait on a cell signal.

Included

Camera set

Six cameras with known optics and mounting hardware. The geometry has to be known to reconstruct from it, which is why the cameras come with the system rather than from a parts bin.

Your choice

Display

Use an iPad you already own, or add a mounted screen. The iPad app is built and reads the same data the wired display does.

// 05. How it works three stages

01 See Six cameras with known positions and lenses produce one common view of the vehicle's surroundings.

02 Reconstruct Depth becomes a grid of occupied and empty space, with a class on each occupied cell. Tracking holds identity on the things that move, so a cyclist behind a parked car is still the same cyclist when they reappear.

03 Show The result is drawn from above and around, on an iPad or a mounted display, with range on everything that matters.

// 06. Measured scene-0103, recorded

Points per frame: 835,613
Render: 13 FPS
Frame budget: 11.3 MS TOTAL
Tracked objects: 25 IN FRAME

// 07. Lane geometry scene-0103 · frame 143

The same scene with lane geometry and the ground plane enabled,
showing painted lane markings running under the reconstructed
point cloud and tracked object boxes. — Lane lines and the ground plane drawn under the reconstruction, with object boxes held through the turn.

// 08. Where this actually stands

No hardware exists yet, and nothing above is a product you can buy. The pictures on this page are the real system running, and what it is running on is recorded street data from a research dataset, not a camera bolted to a van. That is the honest state of it.

What is built: the perception and reconstruction stack, the tracking, the desktop renderer, and a native iPad app reading the same data. What is not: a live camera rig, an enclosure, and everything that turns a working pipeline into something you mount on a vehicle.

The next milestone is running live cameras through it. The system is built so that swapping recorded frames for a live rig touches one layer, which is a design decision rather than a lucky one, but nobody has done it yet.

Tell me when it is real

An email address, nothing else. No payment, no reservation, no delivery date, because promising one would be inventing it. You will hear when a rig runs on a real vehicle.

One address, stored to email you once. Not sold, not shared, not added to anything else.