Skip to content

Work-Zone Cone and Barrier Datasets

Last updated: 2026-05-09

Why It Matters

Temporary traffic-control objects are not just small-object detection targets. Cones, barrels, barricades, barriers, arrow boards, work vehicles, and workers can redefine drivable space, invalidate lane priors, and force temporary map policy changes.

For airside autonomy, the analogous objects are cones, low barriers, chocks, mobile signs, tow bars, staged carts, and temporary closed-area markings. Public roadwork datasets are useful for visual long-tail coverage, but they do not replace apron-specific labels.

Dataset/Benchmark Table

Dataset / sourcePrimary taskSensors / formatUseful labelsBest useMain transfer risk
ROADWorkWork-zone recognition, observation, scene description, and path predictionVehicle-mounted images and video snippets from 18 U.S. cities; COCO-like instance data and pathway data in the public code releaseWork-zone object instances, scene tags, sign/arrow-board attributes, descriptions, road/sidewalk semantic context, 2D/3D traversable pathsTrain and evaluate detectors for cones, drums, barricades, barriers, TTC signs, workers, work vehicles, and scene-level work-zone stateCamera-only road scenes; counts differ between paper and repository releases; no LiDAR/radar or airport object taxonomy
ROADWork discovered subsetWork-zone mining from large driving-image corporaImages mined from BDD100K and Mapillary by ROADWork detector rulesDiscovered roadwork images for auxiliary evaluationTest geographic generalization and open-world work-zone discoveryExternal images inherit source dataset licenses and may not include the full ROADWork annotation stack
RoSARoad construction-zone instance segmentationFHD dashcam video frames from Korean roads, reconstructed into image sequencesArea-level construction-zone masks based on cones/barrels and lane geometryDetect the whole closed or restricted zone rather than only sparse cones or barrelsFocuses on image-space zone masks; paper says a subset will be released and reported frame/split counts need release verification
Existing generic driving datasetsProxy cone/sign appearance and negative examplesMostly camera, sometimes LiDAR depending on datasetSparse cone, sign, lane, vehicle, and pedestrian labelsPretraining, hard-negative mining, and regression testsWork-zone-specific barriers, arrow boards, flaggers, and temporary traffic rules are under-labeled

ROADWork Practical Notes

FieldWhat to capture in an internal benchmark
Object vocabularyPreserve ROADWork-style categories for cones, tubular markers, drums, barricades, barriers, vertical panels, fences, arrow boards, TTC signs, message boards, workers, work vehicles, police actors, and other roadwork objects.
Scene stateLabel whether temporary objects imply an active work zone, a staged inactive zone, or a stray object that should not alter routing.
Sign attributesKeep text and graphic labels separate; use crop-rescale or text-spotting checks when the sign is small.
Pathway targetStore both object masks and a passable path or route decision so the model is not optimized only for pretty segmentation.
Geographic splitSplit by city, airport terminal area, or stand group to expose local cone/barrier designs and deployment habits.

ROADWork is the stronger source for object taxonomy and multimodal scene reasoning. It includes rare categories that open-vocabulary models detect poorly and supports downstream work-zone discovery, sign reading, description, and path prediction.

RoSA Practical Notes

FieldWhat to capture in an internal benchmark
Label unitLabel a continuous construction area, not every cone as a separate planning primitive.
Temporal continuityCarry zone labels through partial occlusion when the previous frames establish the area boundary.
Boundary priorityMeasure the nearest boundary relative to the ego vehicle, because that boundary controls immediate lane availability.
Object exclusionKeep the zone ground mask consistent instead of including irregular above-ground vehicles or workers in the same polygon.
Model baselineRoSA reports YOLOv8 segmentation baselines; use them only as a sanity check, not a final autonomy policy.

RoSA is the better proxy when the operational question is "where is the temporary closed zone?" instead of "which objects are visible?" This maps well to lane closure, stand closure, and temporary exclusion-zone detection.

Metrics

MetricWhat to reportWhy it matters
Instance AP / mask APPer-class AP for cones, barrels, barricades, barriers, signs, workers, and vehiclesRare work-zone classes are easily hidden by aggregate AP
Area IoUIoU of the construction or exclusion-zone polygonPlanning needs the restricted area, not just object centers
Nearest-boundary errorLateral and longitudinal error of the closest closed-zone edgeA small mask error near the vehicle can create an unsafe path
Active-zone precision/recallWhether a scene truly changes road or apron rulesA cone in storage should not trigger rerouting
Sign reading qualityText/graphic recognition accuracy or normalized edit distanceArrow boards and TTC signs can reverse the route decision
Cross-site performanceHold out cities, terminals, contractors, or stand groupsTemporary equipment is highly regional and operator-specific

Airside Transfer

Roadwork conceptAirside analogTransfer useRequired local validation
Cone / barrel lineCone row, stand closure markers, tow-route markersTemporary boundary detection and continuityCone color, height, spacing, night reflectivity, and occlusion by aircraft/GSE
Barrier / barricadeTemporary safety barrier, maintenance cordonClosed-area segmentationLow-profile barriers and non-road layouts on open apron concrete
Arrow board / TTC signMobile sign, marshaller signal board, stand-closure boardAttribute-aware rule changesAirport-specific sign vocabulary and procedures
Worker / police officerGround crew, marshaller, safety escortHuman actor context near temporary zonesPPE, crouched workers, wing/gear occlusion, and low-light operations
Work vehicleMaintenance truck, tug, belt loader, sweeperContext for temporary work stateGSE geometry and parked-vs-active status

Validation Guidance

  1. Use ROADWork first for object vocabulary and rare-category robustness; use RoSA-style labels for the contiguous restricted-area mask.
  2. Build local splits by terminal, contractor, route, and time of day. Random frame splits will overstate generalization.
  3. Evaluate object detection and zone segmentation separately. Missing a cone and misplacing a closed-zone edge are different failures.
  4. Preserve sign/arrow-board text and graphic attributes as structured labels instead of compressing them into one sign class.
  5. Add negative controls: stored cones, equipment yards, parked work trucks, and barriers outside the active route.
  6. For map cleaning, do not let a temporary cone/barrier observation update the permanent static map without multi-session evidence or review.

Sources

Public research notes collected from public sources.