Wireless Vehicular Positioning

• Wireless-based vehicular positioning is a method that uses RF measurements (ToA, AoA, RSSI) to determine vehicle location in challenging, multipath-rich scenarios.
• Cooperative architectures and sensor fusion integrate signals from 5G/6G, LEO satellites, and V2X to meet strict performance targets like sub-meter error and low latency.
• AI and machine learning enhance traditional estimation methods, providing robust, adaptive positioning even in GNSS-denied and high-density traffic conditions.

Wireless-based positioning for vehicular applications encompasses the use of radio-frequency (RF) signals from infrastructure, roadside units, satellites, and other vehicles to determine the state (position, velocity, possibly orientation) of vehicles in dynamic, multipath-rich, and often GNSS-denied environments. Positioning techniques span measurement models (ToA, TDoA, AoA, RSSI, phase), fusion architectures (centralized, decentralized), algorithmic paradigms (model-based estimators, particle/Kalman filters, machine learning, ISAC), and network configurations (infrastructure-centric, cooperative, hybrid). Performance targets are stringent: sub-meter error, <100 ms latency, and >99.99% availability for safety-critical and high-autonomy use cases. Recent advances have integrated cellular (sub-6 GHz, mmWave, THz), V2X sidelink, and LEO constellations with AI/ML and sensor fusion for resilient, scalable, lane-level positioning in urban canyons, tunnels, and high-density traffic regimes.

1. Physical Principles and Measurement Models

Wireless-based vehicular positioning leverages several RF measurement primitives:

• Time-of-Arrival (ToA)/Round-Trip Time (RTT): Absolute or two-way delay estimates from infrastructure (e.g., BS, RSU) to vehicle. For ToA, $d = c\tau$, where $\tau$ is measured delay. RTT eliminates the need for synchronized clocks. Typical ToA errors are inversely proportional to bandwidth and SNR: $$\mathrm{Var}(\tau) \geq \frac{1}{8\pi^2 B \,\mathrm{SNR}}$$ (Ko et al., 2019).
• Time Difference of Arrival (TDoA): Difference in ToA from two or more anchors, forming hyperbolic constraints: $|r - r_A| - |r - r_B| = c\Delta\tau$. Requires multi-anchor network, but cancels absolute clock bias (Dhungel et al., 2024, Fouda et al., 2021).
• Angle-of-Arrival/Departure (AoA/AoD): Multi-antenna arrays estimate direction from phase progression or covariance matrices: $$a(\phi) = [1, e^{j2\pi d \sin\phi / \lambda}, \dots ]^T$$ (Ko et al., 2019, Wymeersch et al., 2019). Variance scales inversely with SNR and array aperture.
• Received Signal Strength (RSS)/Path Loss: $P_{rx} = P_0 - 10\alpha\log_{10}(d) + \eta$ (path-loss exponent $\alpha$, log-normal shadowing $\eta$). Used in coarse multilateration; susceptible to environmental dynamics (Huang et al., 2023, Mohamed et al., 2012).
• Phase-Difference-of-Arrival (PDoA): Exploits cyclic phase wraps over multiple tones for fine-grained delay/range discrimination. Used for both sub-meter ranging at moderate bandwidths and ambiguity resolution in narrowband V2X (Ko et al., 2019).

Measurement noise, clock biases, and multipath/NLoS are dominant error sources. Robust estimators rely on careful error modeling, environment/contextual awareness, and hybridization with motion/perception sensors (Dhungel et al., 2024, Saleh et al., 28 Jan 2026).

2. Cooperative and Decentralized Positioning Algorithms

Cooperative localization exploits inter-vehicle communication and collective measurement to improve absolute and relative positioning.

• Distributed GNSS Augmentation: Double-differenced pseudorange exchanges ($D_{ab}^{i0}$) suppress satellite and receiver biases; vehicles solve small-scale likelihood-constrained multilateration subproblems and fuse outputs via weighted averaging (Liu et al., 2012). Achieves 60–75% error reduction over standalone GNSS.
• Machine Learning-Aided Cooperative Localization (MLCL): Vehicles exchange internal GNSS reads and external ToA/AoA over dynamic V2V graphs, processed via end-to-end DNNs—MTNN (message generation), MRNN (reception/aggregation), SUNN (internal recurrent state), LENN (position estimator). MLCL operates fully decentralized, handles arbitrary topology/link failures, and achieves steady-state MAE ≈3.0 m in dense urban scenarios (vs. 4.2 m for decentralized EKF and >8 m for no-cooperation baselines) (Lee et al., 2024).
• V2X Sidelink RTT/AoA Fusion: RSUs and vehicles form bidirectional RTT/AoA links, enabling clock-bias-free ranging and bearing estimation. Super-resolution parameter extraction (CPD-SA, ESPRIT-SA) achieves 90th-percentile range error ≈0.2 m, azimuth error ≲0.5°, and sub-meter PEB/RMSE across intersection/highway scenes (Ge et al., 2023).
• Backscatter-Tag Assisted Localization: Passive tags at known roadside positions are interrogated by vehicular MIMO-FMCW readers transmitting joint frequency/phase modulated waveforms (JFPM). Multiple-access and broadcast DoFs resolve tag association and vehicle pose; with $N=8$ antennas, ≈0.1 m error is achieved at moderate SNR under realistic mobility (Han et al., 2019).

3. Infrastructure-Based Solutions and Sensor Fusion

Integration with terrestrial and non-terrestrial communication infrastructures is essential for positioning robustness and scalability:

• 5G/6G Cellular Positioning: Downlink-based solutions (NR PRS ToA/TDoA, AoA, multi-band/frequency diversity) are validated on 3GPP-compliant SDR testbeds. Offset calibration for gNB and UE clocks is critical; after correction, sub-2 m RMSE in urban LOS/NLOS is routinely attainable (Dhungel et al., 2024).
• LEO Mega-Constellations: Low Earth Orbit satellites furnish dense, frequently updated line-of-sight (LOS) anchor geometry for lane-level vehicle positioning. Fusing one-way pseudorange, Doppler shift, and TDoA with on-board IMU via WLS/KF/UTC achieves $\lesssim0.1$ m RMSE, even under high mobility and rapid satellite handover. Co-packing with communication/remote sensing boosts accuracy and reduces data-downlink latency (Sheng et al., 2024, Saleh et al., 28 Jan 2026).
• Sensor Fusion Architectures:
  • Loosely coupled: 5G-based absolute position/velocity is fused with IMU/odometer (e.g., LC-EKF/UKF) for high-rate pose tracking. NLOS detection, physically-calibrated process noise, and motion-constraint validation are vital. 14 cm error at 95% coverage is demonstrated in urban trajectories (Saleh et al., 2024, Bader et al., 2023).
  • Tightly coupled: Fusing raw ranges/angles or I/Q with IMU/infrastructure in EKF, UKF, or particle filters achieves sub-decimeter accuracy and rapid NLoS recovery (Hammarberg et al., 2022, Saleh et al., 28 Jan 2026).
• C-V2X/IEEE Standards: RSSI-based multilateration over VANET RSU grids or cellular-V2X (C-V2X) beacons supports robust fallback in GNSS-denied domains (tunnels, urban canyons). Convex SDP estimation and UKF trajectory smoothing in systems such as CV2X-LOCA yield lane-level ($<$4 m) error with spacing $\lesssim$150 m between RSUs (Huang et al., 2023, Mohamed et al., 2012).

4. Machine Learning/AI in Wireless Vehicular Positioning

ML/AI techniques are pervasive for robustness, adaptability, and high-dimensional information extraction:

• Measurement Enhancement: Supervised DNNs/CNNs classify LOS/NLOS, mitigate multipath bias in delay/angle, and denoise CIR/CSI for improved geometric ranging (Pan et al., 24 Jan 2025). Self-/unsupervised learning (e.g., channel charting) preserves spatial topology for fingerprintless localization in fast-changing environments.
• End-to-End Positioning: Neural models (fingerprint-based, channel charting, multi-modal networks) learn direct mappings from radio features (RSSI, RSRP, CSI, ADCPM) to (x,y) position, outperforming classical estimators in deep NLoS and complex urban topologies, often achieving sub-1 m median error at highway speeds (Pan et al., 24 Jan 2025).
• Cooperative Learning: Distributed graph neural networks (GCN, message-passing architectures) jointly process communication and measurement graphs in large vehicular fleets, generalizing well across fleet sizes and link dynamics (Lee et al., 2024).
• AI-driven Fusion and Adaptation: AI modules integrated into standard 3GPP NR framework (Rel-18+) enable dynamic parameter selection, adaptive signal selection, and continuous learning/update for lifelong deployment (Pan et al., 24 Jan 2025).
• ISAC (Integrated Sensing and Communication): Bilinear compressed sensing and GAMP-inspired alternating minimization solve joint demodulation, positioning, and environment imaging using sparse region-of-interest discretization. Simulation results show that sub-meter positioning and robust detection are attainable with moderate complexity (Tong et al., 3 Oct 2025).

5. Challenges, Performance, and Future Directions

Performance Envelope

— Empirically Demonstrated Accuracies:

Technique / System	Typical Error / 90–95%ile	Latency	Scenario	Source
GNSS only (urban)	5–15 m	$\sim$1 s	Open sky (degrades in NLoS/tunnels)	(Saleh et al., 28 Jan 2026)
5G mmWave + INS/ODO (LC UKF)	0.2–0.7 m (avg)	5–20 ms	Dense urban, realistic ray-tracing	(Bader et al., 2023)
MLCL (V2V decentral.)	3.0 m (steady-state MAE)	8 ms/infer	Urban canyons, time-varying dropout	(Lee et al., 2024)
V2X Sidlink (RTT+Aoa+CPD-SA)	0.7 m (RMSE)	100 ms	Urban intersection/highway, 5.9 GHz	(Ge et al., 2023)
C-V2X RSSI + UKF (CV2X-LOCA)	1.5–4 m (ALE)	10 ms	Urban, tunnels, field and sim	(Huang et al., 2023)
LEO GNSS + IMU/KF	0.08–0.4 m (RMSE)	100 ms	Wide-area, rapid handover	(Sheng et al., 2024)
Backscatter-tag (JFPM, N=8)	≈0.1 m	10–100 ms	Roadside tags, LoS	(Han et al., 2019)

Open Research Problems

• Multipath and NLoS Bias: Reliable exploitation (vs. mitigation) of multipath demands robust classification, learning-based filtering, and environment-adaptive radio SLAM (Kim et al., 2019, Bader et al., 2023).
• Anchor/RSU Deployment: Coverage, density, and infrastructural cost remain limiting—optimal spatial/temporal allocation is unsolved under real-world constraints (Huang et al., 2023).
• Synchronization and Scale: Sub-ns time sync and distributed calibration across massive infrastructure and LEO are bottlenecks for sub-meter ToA/TDoA (Dhungel et al., 2024, Sheng et al., 2024).
• Security and Privacy: Message integrity, adversarial RF attacks/spoofing, and privacy-preserving cooperative localization are central for safety assurance (Pan et al., 24 Jan 2025).
• Lifelong Learning/Domain Adaptation: Large-scale, high-label-cost, dynamic environments require continual learning, transfer/meta-learning, and model compression (Pan et al., 24 Jan 2025, Lee et al., 2024).
• End-to-End Data Fusion: Combining cellular, LEO, V2V, UWB, perception, and motion in scalable, integrity-aware frameworks (factor graphs, particle filters, ISAC) is an active area (Hammarberg et al., 2022, Saleh et al., 28 Jan 2026).

6. Historical Context and Standardization Trajectory

Satellite navigation began with Doppler-based TRANSIT (1959) and culminated in multi-constellation GNSS (GPS, GLONASS, Galileo) (Saleh et al., 28 Jan 2026). Cellular positioning evolved from E-CID/AoA in 2G/3G, to OTDoA/PRS/RTK in 4G/5G, and currently toward AI-driven PNT, sidelink, and NTN (Non-Terrestrial Networks) in Rel-18/19/20 of 3GPP. IEEE-based vehicular standards (802.11p, 802.11mc/az, 802.15.4a/z, Bluetooth 5.1/6.0) provide alternative anchor modalities (Wi-Fi ToA/AoA, UWB, BLE AoA) that complement cellular and GNSS coverage (Saleh et al., 28 Jan 2026, Ko et al., 2019).

Emergent 6G/ISAC visions entail joint waveform design for sensing/communication/positioning, RIS-aided virtual LOS creation, and tight integration of LEO, cooperative V2X, and learning-based inference for lane-level, high-availability vehicular localization (Tong et al., 3 Oct 2025, Pan et al., 24 Jan 2025).

---

Wireless-based positioning for vehicular applications is advancing rapidly through cooperative sensor fusion, AI-enhanced measurement, scalable decentralized algorithms, and infrastructure heterogeneity, enabling robust and accurate navigation in the most challenging urban and high-dynamics environments (Lee et al., 2024, Ge et al., 2023, Saleh et al., 28 Jan 2026, Pan et al., 24 Jan 2025).