V2I-GAN: AI-Enhanced Vehicular Communication

• V2I-GAN is a framework that employs generative adversarial networks to enhance vehicular-to-infrastructure communication by integrating Doppler spectrum analysis with dynamic beamforming.
• It combines classical channel models with GAN-based techniques to optimize channel estimation, beam tracking, and waveform design in high mobility and challenging propagation environments.
• The system’s multimodal fusion of CSI and vision data achieves high localization accuracy, supporting smart city applications and autonomous navigation.

Vehicular-to-Infrastructure Generative Adversarial Network (V2I-GAN) encompasses advanced machine learning and signal processing methodologies for optimizing key aspects of communication between vehicles and infrastructure, particularly under high mobility and in challenging propagation environments. V2I-GAN systems integrate classical models of channel dynamics, including Doppler effects in millimeter-wave communications, with generative adversarial networks tasked with enhancing tasks such as channel estimation, waveform design, beam tracking, and multimodal fusion for robust operation in 5G and beyond vehicular networks.

1. Doppler Spectrum in Millimeter-Wave V2I Communications

Millimeter-wave bands, central to 5G and subsequent generations, experience Doppler impairments that scale linearly with carrier frequency, creating pronounced challenges in vehicular environments. Conventional omni-directional reception models, such as Jakes, generate a broad Doppler spectrum due to multipath components impinging over diverse angles. Beamforming fundamentally alters the received Doppler power spectrum by spatially filtering the multipath, so only signals within a preset receive beamwidth ($\theta_{RX}$) are processed.

Key expressions for instantaneous Doppler shift and spectral shaping are as follows:

• Instantaneous Doppler shift:

$f_p = \frac{v}{c}\,f\cos(\theta - \theta_v)$

where $v$ is relative velocity, $c$ speed of light, $f$ carrier frequency, $\theta$ arrival angle, and $\theta_v$ is velocity vector orientation.

• Doppler power spectral density:

$$P(f_a) = \mathrm{rect}\left(\frac{f_a - f_{a,\mathrm{shift}}}{\Delta f_a}\right) \cdot \frac{1}{\Delta f_a}$$

where $$f_{a,\mathrm{shift}} = \frac{f_{D,\max}^+ + f_{D,\max}^-}{2}$$ and $\Delta f_a = f_{D,\max}^+ - f_{D,\max}^-$.

Beamforming confines Doppler spread proportionally to $\theta_{RX}$:

$$\mathrm{Doppler\, spread} \sim f_{D,\max}\cdot\theta_{RX}$$

A direct consequence is suppression of Doppler spread and preservation of a near-pure Doppler shift in the signal, beneficial for robust waveform design and air interface optimization (Lorca et al., 2018).

2. Dynamic Beamwidth Adjustment and Doppler Control

High-speed vehicular scenarios, such as those encountered in railway applications, require mitigation of excessive Doppler spread. Dynamic control of the receiver’s beamwidth is effective for maintaining a constant, manageable Doppler profile. The recommended adjustment guideline:

$$\theta_{RX} (\mathrm{degrees}) = -1.4 \times 10^4\, /\, [f_c(\mathrm{GHz})\cdot v(\mathrm{km/h})]$$

Strategically narrowing $\theta_{RX}$ as velocity increases yields two outcomes: bounded Doppler spread (minimizing inter-carrier interference in OFDM) and increased antenna gain (aiding link margins over larger distances). This enables the designer to maintain subcarrier spacing and system parameters without continuous recalibration for velocity-induced Doppler variation.

3. Integration with Generative Adversarial Networks in V2I Systems

V2I-GAN systems leverage the theoretical Doppler analysis to inform training targets and algorithmic designs for generative adversarial networks addressing:

• Channel state modeling: GAN architectures can learn and predict time-varying channel profiles shaped by beamforming-induced narrow Doppler spectra.
• Beam tracking and inference: GANs exploit predictable Doppler environments to synthesize future beam selection scenarios and refine beam alignment in real-time.
• Interference mitigation: Channel models generated by GANs, informed by physical constraints such as $\theta_{RX}$ and Doppler approximations, are instrumental in proactive resource allocation.

This synthesis of physical theory and data-driven learning constitutes a powerful foundation for AI-enhanced air interface management in V2I deployments.

4. Multimodal Fusion for V2I Localization

Robust V2I operation, especially in urban and GPS-challenged scenarios, relies on accurate localization. Multimodal systems fuse channel state information (CSI) and visual data from infrastructure-equipped vision sensors. The fusion framework employs parallel encoders—a dedicated CSI encoder operating on frequency-domain CSI ($|\hat{H}_F| \in \mathbb{R}^{M \times K}$) and a vision encoder (typically ResNet-18) applied to input image $X_{Vision}$. Feature vectors are projected to a shared, $\ell_2$-normalized embedding space using modality-specific projection heads.

Contrastive loss enforces spatial alignment:

$$\begin{align*} x_{CSI} &= h_{CSI}(z_{CSI}),\quad x_{Vision} = h_{Vision}(z_{Vision}) \ \bar{x}_{CSI} &= x_{CSI}/\|x_{CSI}\|_2,\quad \bar{x}_{Vision} = x_{Vision}/\|x_{Vision}\|_2 \ \mathcal{L}_C &= -\frac{1}{B}\sum_{i=1}^B \log \left(\frac{\exp(S_{ii})}{\sum_{j=1}^B \exp(S_{ij})}\right) \end{align*}$$

Joint embeddings $x_{Joint} = [\bar{x}_{CSI};\,\bar{x}_{Vision}]$ are regressed to produce position estimates $\hat{p}_{UE} = f_{Predict}(x_{Joint})$. Learning minimizes a weighted sum of localization (MSE) and cross-modal alignment losses, adaptively balanced by uncertainty (Zheng et al., 25 Aug 2025).

5. Performance, Error Characteristics, and Computational Considerations

Empirical analysis via the DeepVerse 6G (Urban DT31) dataset demonstrates that multimodal fusion yields localization error as low as 0.55 meters (with 95% of estimates within 1.06 meters), outperforming traditional radio/angle-of-arrival-based estimates and single-modality models (errors exceeding 9 meters). The complementary nature of errors in CSI-only and Vision-only branches is substantiated by near-zero Pearson correlation ($\rho \approx -0.01$), validating the aggregation strategy.

While the fusion framework induces greater computational requirements (parameter count and GFLOPS), modern hardware platforms mitigate overhead, and improvements in accuracy are deemed operationally significant in vehicular networks. This suggests that computational complexity is an acceptable trade-off for the marked increase in localization reliability under varied propagation and visibility conditions.

6. Applications in Modern V2I Environments

The Doppler-management strategies and multimodal GAN-enhanced frameworks are applicable in:

• Autonomous navigation: Ensuring lane-level positioning and safety in GPS-denied urban canyons.
• Smart city infrastructure: Supporting dynamic traffic control, emergency response, and urban mobility analytics.
• Advanced 5G/6G networks: Enabling high-precision localization and resource allocation in infrastructure-supported vehicular communication, informed by adaptive waveform and air interface designs.

A plausible implication is that further advances in GAN model architectures, informed by narrow-band Doppler environments and multimodal sensing, can catalyze development of robust, scalable V2I systems resilient to high mobility and heterogeneous urban propagation.

7. Summary and Future Directions

V2I-GAN systems integrate Doppler spectrum theory (as shaped by beamforming and dynamic beamwidth adaptation) with generative adversarial learning for joint optimization of physical and algorithmic components in vehicular communications. Predictable Doppler spreads, enabled by narrow beamforming, allow for efficient OFDM waveform design and air interface tuning. In parallel, multimodal fusion of CSI and vision data using contrastive learning and regression substantially improves localization accuracy in urban scenarios. These advances position V2I-GAN as a foundational approach for resilient, AI-augmented vehicular-to-infrastructure communication infrastructures in next-generation wireless ecosystems.