V2X Large-Range And Sequential Datasets

Last updated: 2026-05-09

Large-range and sequential V2X datasets extend cooperative perception from isolated frames at a single intersection to multi-agent, multi-sensor, temporally consistent scenes. V2XScenes, V2XPnP Sequential, UrbanIng-V2X, V2X-Real, and related datasets are especially relevant to airport autonomy because fixed infrastructure can observe around occlusions and beyond the ego vehicle's range.

Scope

Dataset	Domain	Main contribution
V2XScenes	Large-range vehicle-infrastructure collaborative perception under challenging traffic conditions	Sequential 3D boxes and tracking IDs across seven roadside layouts and condition-labeled scenes.
V2XPnP Sequential	Real-world multi-agent V2X sequential perception and prediction	Two vehicles, two infrastructure agents, all collaboration modes, trajectories, maps, and benchmarks.
UrbanIng-V2X	Multi-vehicle, multi-infrastructure data across multiple intersections	Larger diversity across intersections with vehicle cameras/LiDAR and infrastructure thermal cameras/LiDAR.
V2X-Real	Real-world multi-modal V2X cooperative perception	2 vehicles + 2 infrastructures, 1.2M annotated 3D boxes, four collaboration modes.
TruckV2X	Synthetic truck-centered cooperative perception	Tractor, trailer, CAV, and RSU agents for heavy-vehicle occlusion and articulation analysis.
CoopScenes	Ego-infrastructure collective perception scenes in Germany	10 Hz synchronized data, multi-scene registration, anonymization, and development kit.
V2X-Seq	Sequential vehicle-infrastructure cooperative perception and forecasting	Sequential perception, trajectories, maps, and forecasting labels across many intersections.

This page is about dataset fit and validation use. It is not a V2X communications-standard page.

Dataset And Task Definitions

Task	What is evaluated
Cooperative 3D object detection	Whether fusing vehicle and infrastructure sensors improves 3D boxes.
Cooperative tracking	Whether object IDs remain consistent across frames and sensor handoffs.
Large-range roadside perception	Whether infrastructure extends range beyond the ego vehicle and through occlusion.
Sequential perception	Whether models use temporal context rather than independent frames.
Prediction / forecasting	Whether trajectories can be forecast from cooperative history and map context.
Collaboration-mode comparison	Vehicle-centric, infrastructure-centric, V2V, V2I, I2I, and full V2X modes.

V2XScenes emphasizes large-range, challenging-condition vehicle-infrastructure perception. V2XPnP emphasizes spatio-temporal fusion for perception and prediction. UrbanIng-V2X emphasizes multi-intersection diversity. V2X-Real emphasizes real-world multi-agent breadth across collaboration modes.

Sensors And Labels

Dataset	Sensors	Labels / metadata
V2XScenes	Mechanical LiDAR, solid-state LiDAR, blind-repair LiDAR, 4D radar, cameras	Large-range sequential 3D boxes, unique tracking IDs, condition descriptions, detection/tracking benchmarks.
V2XPnP Sequential	40K LiDAR frames and 208K camera data from two vehicles and two infrastructure agents	Object trajectories, 136 objects/scene, 10 object types, PCD map, vector map, 100 scenarios, 24 intersections.
UrbanIng-V2X	12 vehicle RGB cameras, 2 vehicle LiDARs, 17 infrastructure thermal cameras, 12 infrastructure LiDARs	Cooperative perception data, HD map/digital twin, OpenCOOD conversion, sequence/frame API.
V2X-Real	Multi-view camera and LiDAR streams from two vehicles and two infrastructure nodes	1.2M annotated 3D boxes, 10 categories, 33K LiDAR frames, 171K camera data, four sub-datasets.
TruckV2X	LiDAR and RGB cameras across tractor, trailer, CAV, and RSU agents	64 synthetic scenarios, 88,396 LiDAR frames, 1.18M 3D boxes, occlusion recovery analysis, heavy/light/VRU slices.
CoopScenes	Ego vehicle six-camera/three-LiDAR setup plus infrastructure towers	104 minutes at 10 Hz, 62K frames, synchronized/registered/anonymized scenes.
V2X-Seq	Vehicle and infrastructure sequential data	Sequential perception frames, trajectories, vector maps, traffic lights, forecasting scenarios.

The most useful datasets for airside design are those with fixed infrastructure and sequence-level identity, because aircraft stands and apron roads are dominated by occlusion, long-range approach, and handoff between viewpoints.

Metrics

Metric	Use
3D mAP / AP by IoU	Detection quality for cooperative boxes.
NDS-style detection metrics	Balanced detection score where used by the benchmark.
AMOTA / AMOTP or tracking-metric variants	Tracking accuracy and localization over time.
ID switches / fragmentation	Whether cooperative handoff breaks object identity.
Range-sliced AP	Measures infrastructure value at long range.
Occlusion-sliced AP	Measures whether cooperation actually solves blind spots.
Mode delta	Difference between vehicle-only, infrastructure-only, V2V, V2I, I2I, and full fusion.
Latency/bandwidth-aware score	Needed when transmitted features or boxes have deployment constraints.

For airside work, add clearance-specific metrics: missed actor inside the swept path, missed actor under aircraft envelope, handoff delay at stand entry, and false-clear in occluded pushback zones.

Failure Modes

Calibration errors between fixed infrastructure and vehicle sensors create ghost boxes or shifted tracks.
Time synchronization errors are amplified in multi-agent fusion, especially for moving vehicles and pedestrians.
Infrastructure-heavy datasets can overstate performance if the deployment cannot install equivalent sensor coverage.
Sequential labels may hide dropped-frame or packet-delay behavior unless communication is simulated separately.
Large-range detection may improve vehicle recall while still missing low-profile objects close to the ego vehicle.
Tracking IDs can break when an object moves between infrastructure fields of view.
Public-road V2X classes do not cover aircraft, tugs, belt loaders, dollies, ULDs, cones, chocks, hoses, and FOD.
Weather and night coverage varies widely by dataset and should not be assumed from "real-world" alone.

AV, Indoor, Outdoor, And Airside Relevance

Environment	Fit	Notes
Public-road AV	Strong	These are built for cooperative driving perception and prediction.
Airport airside	Strong architecture proxy	Fixed infrastructure, controlled network, and repeated routes make V2X especially practical.
Indoor logistics	Moderate	Multi-agent and fixed-camera principles transfer, but sensors/maps differ.
Outdoor industrial sites	Strong	Similar intersections, yards, and occluded equipment corridors.
Long-range safety monitoring	Strong	Infrastructure can observe beyond ego line of sight.

Airports may be easier than public roads in one respect: the operator can decide where to place infrastructure sensors and can mandate network participation. They are harder in another respect: aircraft geometry, stand equipment, GSE, reflective surfaces, and operating rules are outside public-road datasets.

Validation And Data-Engine Use

Use V2X-Real and V2XPnP to compare collaboration modes before designing an airport sensor topology.
Use V2XScenes to study large-range handoff, condition labels, and cooperative tracking under challenging traffic.
Use UrbanIng-V2X and CoopScenes to stress multi-site calibration, synchronization, and OpenCOOD-style data conversion.
Use TruckV2X when the target fleet includes articulated trucks, trailers, long GSE, or other large occluders whose geometry is missing from light-vehicle V2X datasets.
Build an airport dataset with fixed stand cameras/LiDAR/radar, vehicle sensors, synchronized clocks, maps, and unique IDs across stands.
Keep vehicle-only, infrastructure-only, and fused outputs in logs; the data engine needs to know which source found each object.
Mine disagreement cases: infrastructure sees an object the vehicle misses, vehicle sees an object infrastructure misses, and fusion suppresses a correct single-agent detection.
Add communication replay with latency, packet drop, stale transforms, and degraded collaborator trust before using V2X outputs in a safety argument.

SLAM Methods

Methods

V2X Large-Range And Sequential Datasets ​

Scope ​

Dataset And Task Definitions ​

Sensors And Labels ​

Metrics ​

Failure Modes ​

AV, Indoor, Outdoor, And Airside Relevance ​

Validation And Data-Engine Use ​

Sources ​