Left / Right arrows or Page Up / Down. Home / End. Keys 1–6. P to print. No network required.

Vehicle localization from the cameras you already runSlide 1 of 6. TRACKGUARD Vehicle localization from the cameras you already run Seven calibrated cameras, a reference map and motion signals produce one shared vehicle pose. When geometric evidence is insufficient, the output is unavailable. REAL TRACK / SEALED DEVELOPMENT REPLAY Laps 6 + 7. One seed. 49 unavailable outputs. POSE AVAILABLE x y heading UNAVAILABLE No pose published TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026 01TRACKGUARDVehicle localizationfrom the camerasyou already runSeven calibrated cameras, a reference map andmotion signals produce one shared vehicle pose.When geometric evidence is insufficient, the outputis unavailable.REAL TRACK / SEALED DEVELOPMENT REPLAYMeasured R16 track replay with position and yaw error. Reference is evaluation-only; missing outputs remain gaps.Laps 6 + 7. One seed. 49 unavailable outputs.POSE AVAILABLEx y headingUNAVAILABLENo pose publishedTRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 202601
The customer decision: evaluate an additional position signalSlide 2 of 6. CUSTOMER DECISION An additional position signal for a mapped course Vehicle autonomy leads and track operators When the existing position signal weakens, can camera evidence provide a useful second input? THE PILOT DECISION Is the signal accurate and available often enough for our course and operating conditions? Existing inputs Compatible calibrated cameras and a reference map Proposed integration A timestamped pose and explicit availability status Value to test Useful supplementary coverage within the existing stack TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026 02CUSTOMER DECISIONAn additional position signal for a mapped courseVehicle autonomy leadsand track operatorsWhen the existing position signalweakens, can camera evidence providea useful second input?THE PILOT DECISIONIs the signal accurate and available oftenenough for our course and operatingconditions?Existing inputsCompatible calibrated camerasand a reference mapProposed integrationA timestamped pose and explicitavailability statusValue to testUseful supplementary coveragewithin the existing stackTRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 202602
One vehicle pose from map geometry, cameras and motionSlide 3 of 6. SOLUTION One vehicle pose from map geometry and motion SYNTHETIC ILLUSTRATIVE GEOMETRY Reference map Map correspondences constrain camera poses. BODY 7 calibrated cameras 1 shared body pose Calibration ties cameras to one vehicle frame. The quality threshold remains uncalibrated. CAUSAL ESTIMATION 01 Geometry Match the map and solve a consistent body pose. 02 Motion Fuse motion signals with current and past evidence. 03 Quality gate Publish an accepted pose or mark it unavailable. TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026 03SOLUTIONOne vehicle pose from map geometry and motionSYNTHETIC ILLUSTRATIVE GEOMETRYReference mapMap correspondencesconstrain camera poses.BODY7 calibrated cameras1 shared body poseCalibration ties cameras to one vehicle frame.The quality threshold remains uncalibrated.CAUSAL ESTIMATION01 GeometryMatch the map and solve a consistent body pose.02 MotionFuse motion signals with current and past evidence.03 Quality gatePublish an accepted pose or mark it unavailable.TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 202603
Same-session development meets aggregate targetsSlide 4 of 6. VERIFIED AGGREGATE EVIDENCE Same-session development meets aggregate targets Laps 6 + 7 / One continuous run / One initial truth seed / No lap 7 reset 17.48cm Position RMSE Target ≤20cm 11.42cm Lateral RMSE Target ≤20cm 94.8% One-second coverage Target ≥90% Coverage counts occupied one-second bins. Development data is consumed, not untouched validation. TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026 04VERIFIED AGGREGATE EVIDENCESame-session development meets aggregate targetsLaps 6 + 7 / One continuous run / One initial truth seed / No lap 7 reset17.48cmPosition RMSETarget ≤20cm11.42cmLateral RMSETarget ≤20cm94.8%One-second coverageTarget ≥90%MetricObservedHow to read itPosition error, p9531.95cmRMSE is not a per-output guaranteeYaw / longitudinal RMSE0.107° / 13.23cmWrapped heading / along-track positionWithin 20cm, all planned71.24%327/459 non-seed frames, unavailable = missPublished / unavailable411 / 49460 planned, 89.35% publication fractionCoverage counts occupied one-second bins. Development data is consumed, not untouched validation.TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 202604
Coverage and cold start remain the deployment blockersSlide 5 of 6. OPERATIONAL BOUNDARIES Coverage and cold start remain the deployment blockers Original 131-second stress window 18.48 / 9.95cm Position / lateral RMSE on published poses 72% coverage: FAIL 179/263 published, 84 unavailable. Separate seeded window. Below 90% target. Blind day replay 0 / 148 Published / planned groups, all unavailable Never initialized No truth labels or initial seed. No RMSE. No approved three-camera fix. Next proof: independent day/night runs, reliable cold starts and calibrated confidence. End-to-end latency is unproven. No production deployment or safety certification. TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026 05OPERATIONAL BOUNDARIESCoverage and cold start remainthe deployment blockersOriginal 131-second stress window18.48 / 9.95cmPosition / lateral RMSE on published poses72% coverage: FAIL179/263 published, 84 unavailable.Separate seeded window. Below 90% target.Blind day replay0 / 148Published / planned groups, all unavailableNever initializedNo truth labels or initial seed. No RMSE.No approved three-camera fix.Next proof: independent day/night runs, reliable cold starts and calibrated confidence.End-to-end latency is unproven. No production deployment or safety certification.TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 202605
A scoped pilot to decide whether this signal earns integrationSlide 6 of 6. PROPOSED CUSTOMER PILOT One vehicle. One mapped course. A decision in four weeks. Offline replay, then shadow operation Use compatible existing cameras and recorded motion signals. Keep the output outside vehicle control. Customer contribution Course access, calibration, an independent evaluation capture and an agreed latency budget. Decision: extend the pilot only if the agreed gates pass. TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026 06PROPOSED CUSTOMER PILOTOne vehicle. One mapped course. A decision in four weeks.Offline replay, thenshadow operationUse compatible existing camerasand recorded motion signals.Keep the output outside vehiclecontrol.Customer contributionCourse access, calibration, anindependent evaluation capture andan agreed latency budget.Acceptance itemProposed targetIndependent day + nightFreeze map and parameters before capturePosition / lateral RMSE≤20cm each on published non-seed posesOne-second coverage≥90% in each agreed test conditionCold start, no truth seed≥95% of agreed starts within 30sTail error + missing posesReport p95 and all-planned within 20cmSignal integrity + latency100% explicit unavailable states, measure p95Decision: extend the pilot only if the agreed gates pass.TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 202606

Slide text and evidence notes

Slide 1: Vehicle localization from the cameras you already run

TRACKGUARD

Vehicle localization
from the cameras
you already run

Seven calibrated cameras, a reference map and motion signals produce one shared vehicle pose.

When geometric evidence is insufficient, the output is unavailable.

REAL TRACK / SEALED DEVELOPMENT REPLAY

Laps 6 + 7. One seed. 49 unavailable outputs.

POSE AVAILABLE

x y heading

UNAVAILABLE

No pose published

TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026

01

Source: data/metrics.json and data/replay.json (approved derived evaluation display, 2026-10-22). The cover shows the real sealed R16 development replay. Architecture diagrams are illustrative. Raw recordings, world-coordinate poses and labels remain local. Recorded real-track replay of the selected R16 run; its reference is evaluator-only after sealed inference. Animation is accelerated and does not measure live throughput. The interactive dashboard retains all 460 rows.

Slide 2: The customer decision: evaluate an additional position signal

CUSTOMER DECISION

An additional position signal for a mapped course

Vehicle autonomy leads
and track operators

When the existing position signal weakens, can camera evidence provide a useful second input?

THE PILOT DECISION

Is the signal accurate and available often enough for our course and operating conditions?

Existing inputs

Compatible calibrated cameras and a reference map

Proposed integration

A timestamped pose and explicit availability status

Value to test

Useful supplementary coverage within the existing stack

TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026

02

Source: data/metrics.json and data/replay.json (approved derived evaluation display, 2026-10-22). The cover shows the real sealed R16 development replay. Architecture diagrams are illustrative. Raw recordings, world-coordinate poses and labels remain local. Customer value and integration fit are hypotheses for a pilot. No customer deployment, ROI or hardware saving has been measured. This is a supplementary signal, with no actuator authority in the proposed pilot.

Slide 3: One vehicle pose from map geometry, cameras and motion

SOLUTION

One vehicle pose from map geometry and motion

SYNTHETIC ILLUSTRATIVE GEOMETRY

Reference map

Map correspondences constrain camera poses.

BODY

7 calibrated cameras
1 shared body pose

Calibration ties cameras to one vehicle frame.
The quality threshold remains uncalibrated.

CAUSAL ESTIMATION

01 Geometry

Match the map and solve a consistent body pose.

02 Motion

Fuse motion signals with current and past evidence.

03 Quality gate

Publish an accepted pose or mark it unavailable.

TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026

03

Source: data/metrics.json and data/replay.json (approved derived evaluation display, 2026-10-22). The cover shows the real sealed R16 development replay. Architecture diagrams are illustrative. Raw recordings, world-coordinate poses and labels remain local. Architecture is a conceptual explanation from the task and design contract. It does not specify unsupported sensor rates or production interfaces. Publication uncertainty threshold is uncalibrated, not an actual-error bound.

Slide 4: Same-session development meets aggregate targets

VERIFIED AGGREGATE EVIDENCE

Same-session development meets aggregate targets

Laps 6 + 7 / One continuous run / One initial truth seed / No lap 7 reset

17.48cm

Position RMSE

Target ≤20cm

11.42cm

Lateral RMSE

Target ≤20cm

94.8%

One-second coverage

Target ≥90%

Coverage counts occupied one-second bins. Development data is consumed, not untouched validation.

TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026

04

MetricObservedHow to read it
Position error, p9531.95cmRMSE is not a per-output guarantee
Yaw / longitudinal RMSE0.107° / 13.23cmWrapped heading / along-track position
Within 20cm, all planned71.24%327/459 non-seed frames, unavailable = miss
Published / unavailable411 / 49460 planned, 89.35% publication fraction

Source: data/metrics.json and data/replay.json (approved derived evaluation display, 2026-10-22). The cover shows the real sealed R16 development replay. Architecture diagrams are illustrative. Raw recordings, world-coordinate poses and labels remain local. Position RMSE sqrt(mean(dx^2+dy^2)) on finite published estimates, excluding initial truth seed. Lateral error projects across ground-truth heading. One-second coverage differs from publication fraction. {"id":"later_laps","label":"Development laps 6 + 7","kind":"same-session development","coverage_percent":94.8,"planned":460,"published":411,"unavailable":49,"seed_count":1,"qualified":true,"note":"One continuous run. One initial truth seed. No reset between laps. Consumed development data, not untouched validation.","position_rmse_m":0.1747753326636938,"lateral_rmse_m":0.11421952782837566,"position_median_m":0.13351518078480298,"position_p95_m":0.31945388599089014,"position_max_m":0.5874495693389843,"within20cm_all_planned_percent":71.24183006535948,"within20cm_count":327,"scored_planned_excluding_seed":459,"paired_scored_frames":410,"longitudinal_rmse_m":0.13228876131541847,"yaw_rmse_deg":0.10694967979624358,"absolute_lateral_p95_m":0.25184152755336653,"absolute_lateral_max_m":0.5706257251895852,"absolute_longitudinal_p95_m":0.2595835414358725,"absolute_longitudinal_max_m":0.39985447463360063,"absolute_yaw_p95_deg":0.21044725023463787,"absolute_yaw_max_deg":0.41764192530648103}

Slide 5: Coverage and cold start remain the deployment blockers

OPERATIONAL BOUNDARIES

Coverage and cold start remain
the deployment blockers

Original 131-second stress window

18.48 / 9.95cm

Position / lateral RMSE on published poses

72% coverage: FAIL

179/263 published, 84 unavailable.
Separate seeded window. Below 90% target.

Blind day replay

0 / 148

Published / planned groups, all unavailable

Never initialized

No truth labels or initial seed. No RMSE.
No approved three-camera fix.

Next proof: independent day/night runs, reliable cold starts and calibrated confidence.
End-to-end latency is unproven. No production deployment or safety certification.

TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026

05

Source: data/metrics.json and data/replay.json (approved derived evaluation display, 2026-10-22). The cover shows the real sealed R16 development replay. Architecture diagrams are illustrative. Raw recordings, world-coordinate poses and labels remain local. Stress is not ranked. Day: 73.9s, 148 complete groups at 2 Hz, 145 geometry rejects, 3 single-camera results. Day coverage uses 74 camera-time bins without ground-truth interval. No independent-session 20cm result, live 10Hz measurement, GPU speed result, deployment or safety certification. Nominal 0.20m publication uncertainty threshold is uncalibrated and cannot guarantee actual error. Earlier R14 raw geometry timing is not R16 end-to-end latency.

Slide 6: A scoped pilot to decide whether this signal earns integration

PROPOSED CUSTOMER PILOT

One vehicle. One mapped course. A decision in four weeks.

Offline replay, then
shadow operation

Use compatible existing cameras and recorded motion signals. Keep the output outside vehicle control.

Customer contribution

Course access, calibration, an independent evaluation capture and an agreed latency budget.

Decision: extend the pilot only if the agreed gates pass.

TRACKGUARD / CUSTOMER PILOT BRIEF / 22 OCT 2026

06

Acceptance itemProposed target
Independent day + nightFreeze map and parameters before capture
Position / lateral RMSE≤20cm each on published non-seed poses
One-second coverage≥90% in each agreed test condition
Cold start, no truth seed≥95% of agreed starts within 30s
Tail error + missing posesReport p95 and all-planned within 20cm
Signal integrity + latency100% explicit unavailable states, measure p95

Source: data/metrics.json and data/replay.json (approved derived evaluation display, 2026-10-22). The cover shows the real sealed R16 development replay. Architecture diagrams are illustrative. Raw recordings, world-coordinate poses and labels remain local. Proposed pilot, scope and acceptance targets, not measured achievements or committed delivery dates. Week 1 calibrate/freeze/evaluation design, week 2 independent capture, week 3 replay and cold-start evaluation, week 4 shadow integration only if earlier gates pass. Agree denominator and trial count for cold-start target before capture. No truth seed enters the proposed deployment signal. Independent ground truth is evaluation only and must remain separate from map construction. Timing target requires customer stack budget agreed before evaluation. Report failures and unavailable outputs. No safety certification or ROI claim.