Camera-based vehicle localization

Vehicle localization from the cameras you already run.

Position and heading from seven calibrated cameras, a reference map, vehicle speed and yaw rate. Weak geometric evidence produces an explicit unavailable output.

Explore the evidence Read the approach

A localization signal for your existing stack. Deployment readiness and hardware savings remain pilot hypotheses.

Measured development replayR16 · laps 6 + 7

Measured replay not loaded

Replay paused · data not loadedInspect measured replay ↓

Evaluation-only reference. No future reference enters inference. Overview scale cannot resolve centimetre errors.

07 calibrated camerasReference mapVehicle speed and yaw rateVehicle pose or unavailable

Recorded trajectory

Measured development replay

View replay JSON ↗

R16 continuous run, laps 6 + 7. Same-session, seeded development evidence. This replay is separate from the five aggregate windows below.

Measured replay not loaded. Controls become available when the approved data loads.

Real track image not loaded

Evaluation-only referencePublished estimate

Real track bounds and recorded reference path. Rounded display pixels only. Gaps remain unavailable; error charts resolve detail the overview cannot show.

Current frameNot loaded

Frame unavailable · lap unavailable

Position error · absolute: Unavailable
Yaw error · signed, wrapped: Unavailable
Lateral error · signed: Unavailable
Longitudinal error · signed: Unavailable
10 s rolling position RMSE: Unavailable
10 s rolling frame publication: Unavailable

Rolling RMSE uses eligible published non-seed observations. Rolling publication counts all planned frames, including unavailable outputs. It is separate from one-second coverage.

Speed 0:00.0 / unavailablePaused

Slider arrows seek one frame. Space plays or pauses on the slider or track. Click a chart to seek.

Absolute position error

cmUnavailable

Measured data not loaded

Signed yaw error

wrapped degreesUnavailable

Measured data not loaded

Signed directional errors

Lateral Longitudinal · cmUnavailable

Measured data not loaded

Shared cursor: elapsed recording time in seconds. Missing errors and the unscored initial seed are gaps, never zero. Yaw is wrapped to ±180°. No interpolation across missing outputs.

Measured evidence

Precision is one part of the result.

Local targets ≤ 20 cm RMSE≥ 90% one-second coverage

Development onlySame-session, seeded evidence

Stress: coverage failsNot ranked

Day: no initializationRMSE unavailable

Aggregate window only

Position RMSE

Lateral RMSE

Occupied one-second coverage

Frame publication fraction

Longitudinal RMSE

Unavailable

Yaw RMSE · wrapped error

Unavailable

Aggregates for the selected window. Replay controls above remain on the continuous laps 6 + 7 run.

RMSE is not a per-output guarantee.

Unavailable frames count as misses in the within-20-cm share.

Position error p95

Planned frames within 20 cm

Accuracy against the target

RMSE in centimetres · lower is better

Only finite published estimates contribute to RMSE. Initial truth seed excluded.

Availability has two denominators

Occupied time bins and published frames

Compare every window

Overlapping windows. Do not sum.

All five evaluation windows with local RMSE targets of 20 cm and occupied one-second coverage target of 90 percent. Select a window to inspect it.
Evaluation window	Position RMSE	Lateral RMSE	1 s coverage	Published / planned	Truth seeds	Local result

Compute context

Raw geometry profile.

Mean

p95

Live performance unproven

Scope and definitions

Know what the numbers mean.

View aggregate JSON

How accuracy, availability and initialization are evaluated

Exact values for the selected window

Unrounded approved aggregates. Presentation values above are rounded.

Approved embedded aggregate snapshot