Vehicle-to-Infrastructure Measurement Campaign

• Vehicle-to-Infrastructure measurement campaigns are systematic efforts that capture wireless and sensor data between vehicles and roadside infrastructure using diverse drive-tests and multi-modal sensing.
• Advanced instrumentation and synchronization techniques, including SDR logging and sub-millisecond calibration, ensure high-precision data acquisition for ITS and C-V2X systems.
• Empirical findings on metrics like RSRP, delay spread, and Doppler shifts guide channel modeling and inform robust design and optimization of cooperative perception systems.

Vehicle-to-Infrastructure (V2I) Measurement Campaigns constitute systematic efforts to characterize the physical, protocol, and perception properties of wireless and sensing links between vehicles and roadside infrastructure in environments ranging from dense urban centers to industrial halls. These campaigns establish empirical benchmarks for link-quality, cooperative localization, channel modeling, and multi-agent situational awareness. Methodologies span drive-test radio surveying, multi-modal sensor fusion, controlled site deployments, and machine learning dataset construction. V2I campaigns serve as the empirical basis for the design, calibration, and validation of modern ITS, C-V2X, and cooperative perception systems.

1. Experimental Architectures and Environments

V2I measurement campaigns leverage diverse architectural layouts, sensor suites, and deployment Topologies. Urban drive-tests frequently utilize software-defined radios (SDRs) as vehicular probes (e.g., Ettus USRP-N210 with GPSDO) in commercial LTE environments (Vasudeva et al., 2019). Campaigns may span long-range urban-rural corridors (e.g., Miami–Raleigh, 1300 km, LTE carrier fixed at 739 MHz) or be localized to representative city segments (Berlin, 17.2 km loop, multiple drive formations) (Hernangómez et al., 2022). Multi-modal datasets collect RGB/thermal cameras, LiDARs, mmWave radars, ultra-wideband (UWB), IMUs, GNSS-RTK, and occupancy grids, both on vehicles and roadside sensor poles (Qin et al., 23 Dec 2025, Sekaran et al., 27 Oct 2025). Infrastructure installations vary: poles mounted 5–8 m above intersections for field-of-view maximization; RSUs at regular 400–600 m intervals along dense urban loops (Qin et al., 23 Dec 2025, Sekaran et al., 27 Oct 2025).

Industrial measurement campaigns deploy Automated Guided Vehicles (AGVs) with private LTE modems and full navigation stacks in factory halls with metallic obstacles, both LOS and NLOS segments, and long driving times (≈16 h over multiple days) (Hernangómez et al., 2022). mmWave campaigns exploit static RSU–vehicle pairs with tightly synchronized radar and phased array setups at 24 GHz and 60 GHz (Ali et al., 2019, Groll et al., 2019).

2. Instrumentation, Synchronization, and Calibration

All campaigns synchronize vehicle and infrastructure data to sub-millisecond or sub-microsecond accuracy using PPS-aided PTP (Qin et al., 23 Dec 2025), GPS receivers, or NTP. ROS-bag and CSV/Parquet datafiles are timestamped to enable multi-sensor fusion. Calibration workflows encompass intrinsic calibration (checkerboard methods for cameras; factory-characterized LiDAR/radar), extrinsic calibration (rigid-body SE(3) transformations determined via ArUco boards, ICP, total-station measurements) (Qin et al., 23 Dec 2025, Sekaran et al., 27 Oct 2025, Hernangómez et al., 2022), and temporal alignment (phase-locking, hardware-triggered scan start). Data fusion protocols employ motion compensation for intra-scan point cloud alignment and global frame projection (ENU) for multi-agent cooperative perception (Sekaran et al., 27 Oct 2025).

In radio campaigns, SDRs log full I/Q streams, protocol traces, and physical layer counters at multi-millisecond intervals. Packet-level traces and PHY/MAC/RRC events are logged via tools such as MobileInsight (Hernangómez et al., 2022, Hernangómez et al., 2022). Delay-spread and Doppler metrics demand time–frequency domain sounders with sub-μs resolution (Groll et al., 2019).

3. Data Collection, Post-Processing, and Feature Engineering

Data collection spans geo-tagged wireless measurements (SNR, RSRP, RSSI, RSRQ), protocol-level events (cell ID transitions, handovers), environment attributes (area_label, traffic, weather), and physical traces (vehicle speed, heading, pose). Sampling intervals range from 10 ms (PHY counters) to 1 s (IP throughput/jitter, GPS), with map-matching for spatial labeling (Hernangómez et al., 2022).

Post-processing includes OFDM synchronization, frequency offset compensation (CP phase estimation), cell search over PSS/SSS pairs, PBCH decoding for MIB/SIB1 extraction, and reference signal power calculation (Vasudeva et al., 2019). Feature extraction for ML employs sliding-window statistics (mean/std/min/max of PHY metrics), temporal autocorrelation, cross-feature ratios, and contextual environmental variables (Hernangómez et al., 2022).

Delay–Doppler campaigns compute the local scattering function (LSF) via symplectic FFT, DPSS multitapers, and sparse Bayesian learning to extract DD-bins and multipath support points (Groll et al., 2019). mmWave beam-training logs SNR matrices over codebook sweeps; radar APS guides restricted search to high-probability beam pairs (Ali et al., 2019).

4. Channel, Link, and Navigation Metrics

Typical Key Performance Indicators (KPIs) and channel metrics include:

Metric	Definition/Method	Reported Ranges
RSRP	Mean received ref signal power per RE	–75 dBm (urban), –85 dBm (rural) (Vasudeva et al., 2019)
Outage Probability	Fraction of time RSRP < –100 dBm	Urban: 5%; Rural: 25% (Vasudeva et al., 2019)
Delay Spread (τ_rms)	sqrt of second moment of power-delay profile	50–200 ns (factory to multi-path) (Hernangómez et al., 2022), <20 ns (mmWave urban canyon) (Groll et al., 2019)
Doppler Spread (Δν_rms)	max Doppler over all rays	up to 2.8 kHz (50 km/h car) (Groll et al., 2019)
Packet Error Ratio (PER)	Sidelink error fraction vs distance	80 m (CAM), 30 m (CPM) (Hernangómez et al., 2022)
GNSS-RTK RMSE	Absolute position error	6.8–18.6 m (urban loop), 8.15 m (RSU base) (Qin et al., 23 Dec 2025)
Cooperative Perception mAP	3D box average precision @ IoU 0.5	drop of 0.14 mAP for unseen intersection splits (Sekaran et al., 27 Oct 2025)

Link-level benchmarks clarify that urban deployments—denser cells, small-cells, and low-frequency bands—produce stronger coverage, lower outage probability, and sub-1 s disconnect durations (Vasudeva et al., 2019, Hernangómez et al., 2022). mmWave regimes demonstrate extreme DD-sparsity, mainly single-LOS clusters and few strong multipaths, supporting low-complexity beam-tracking and OTFS modulation (Groll et al., 2019). Sidelink communication achieves reliable V2I range up to 80 m for CAM; reduced to 30 m for higher-frequency CPM (Hernangómez et al., 2022).

5. Statistical Findings and Channel Modeling

Empirical channel models fit log-distance path loss and shadow fading terms for RSRP distributions across environment types:

$$PL(d) = PL(d_0) + 10\,\gamma \log_{10}\left(\frac{d}{d_0}\right) + X_\sigma$$

Observed path-loss exponents and shadowing parameters may be derived from dataset regression (Vasudeva et al., 2019, Hernangómez et al., 2022). Urban V2I links cluster around 2 km mean distance (RSRP > –80 dBm); rural environments extend to mean 4 km, outage rates up to 25%. Velocity–RSRP joint PDFs confirm increased fading variance at higher speeds (~120 km/h) (Vasudeva et al., 2019).

Delay–Doppler domain measurements in urban mmWave (60 GHz) street canyons confirm >80% power concentration in one LOS-dependent cluster, with 5–7 significant secondary peaks; delay-spread Δτ_rms ≲20 ns, Doppler-spread Δν_rms ≲300 Hz. The 15° up-tilt beam shows ~10–15% reduction in delay-spread (Groll et al., 2019).

Radar-guided beam-training for mmWave V2I slashes search complexity (96.9% reduction in training time) and boosts beam recovery compared to GNSS-only aiding, especially in NLOS regimes (Ali et al., 2019). Path-loss measurements for lower LTE bands indicate additional 30–50 dBm coverage in challenging urban or rural propagation settings (Vasudeva et al., 2019).

6. Dataset Annotation, Structure, and Evaluation Protocols

Datasets released from Berlin V2X (Hernangómez et al., 2022), UrbanV2X (Qin et al., 23 Dec 2025), UrbanIng-V2X (Sekaran et al., 27 Oct 2025), and iV2I+ (Hernangómez et al., 2022) cover labeled environment classes (e.g., residential, highway, tunnel), annotated sensor streams, and metadata for ML onboarding. Sensor-frame transformations conform to SE(3) conventions; all timebases are unified to UTC/PTP. Annotations feature 3D bounding boxes (at 10 Hz), per-object class, occlusion flags, and trajectory tags (Sekaran et al., 27 Oct 2025). Dataset file trees hold calibration YAMLs, pose ground truth, and parser scripts for rapid inclusion in model pipelines (Qin et al., 23 Dec 2025).

Evaluation protocols encompass train/val/test splits by intersection, time, or scenario; fusion strategies (early/late/intermediate); and per-class average precision at IoU thresholds. Cross-split results elucidate generalization and overfitting effects in cooperative perception (Sekaran et al., 27 Oct 2025).

7. Implications for V2I System Design and Research Directions

Measurement results directly inform design parameters for C-V2X, ITS, industrial AGV networks, and cooperative perception. Cell-site density should be limited to ≈3 km in rural highways and 0.5–1 km in urban cores to guarantee sub-1 s outage durations and low-latency handover (<100 ms) (Vasudeva et al., 2019). Fusion of infrastructure sensor outputs, including ray-based radar APS for mmWave beam selection, minimizes training overhead and enhances link reliability (Ali et al., 2019). Sidelink V2I and cellular V2N trace datasets support transfer learning, sojourn time prediction, route-aware beamforming, and ML-based QoS forecasting across diverse RATs, operators, and sensor stacks (Hernangómez et al., 2022).

Future campaigns should extend spatial and temporal sampling, incorporate more varied environments (narrow alleys, open highways, complex intersections), and deploy high-density RSUs at optimal heights for LoS/NLoS coverage (Qin et al., 23 Dec 2025). Dataset structure and labeling conventions are essential for reproducibility and benchmarking of federated, multi-agent ML systems.

A plausible implication is that the increasing resolution and temporal fidelity of V2I measurement datasets will enable not only channel modeling but also real-time predictive control and context-aware optimization of urban mobility ecosystems.