Reading Radio from Camera: Visually-Grounded, Lightweight, and Interpretable RSSI Prediction

Technical Summary: Reading Radio from Camera

Problem Statement
Accurate, real-time prediction of the Received Signal Strength Indicator (RSSI) is critical for the design and optimization of next-generation wireless networks. Traditional analytical models require precise environmental parameters and complex ray tracing, often limiting their use to simulations. Conversely, existing deep learning approaches that predict RSSI directly from visual inputs (e.g., RGB images) typically function as "black boxes." These brute-force, end-to-end regression models lack physical interpretability, struggle with environmental interference, and often rely on computationally intensive architectures unsuitable for resource-constrained edge devices.

Methodology
The authors propose a novel, physics-guided framework that decomposes the complex task of RSSI prediction into simpler, physically interpretable sub-problems. Instead of predicting the final RSSI value directly, the model predicts two fundamental components of wireless propagation:

1. Path Loss (PL): Distance-based attenuation.
2. Shadow Fading (SH): Fluctuations caused by obstacles obstructing the signal path.

The final RSSI is reconstructed using the physical relationship: $RSSI = -PL + SH$ (assuming a stable transmit power proxy).

Architecture and Inputs
The framework utilizes a lightweight, multi-branch architecture based on MobileNet for visual efficiency. It processes three distinct input modalities:

• RGB Images: Captured from the receiver's perspective.
• 2D Relative Position: A vector $(dx, dy)$ derived from GPS coordinates.
• Bounding Boxes: Information regarding the transmitter and up to 10 potential distractors (e.g., vehicles, pedestrians).

The model employs three parallel encoders:

• Image Encoder: Extracts high-level semantic features using MobileNet.
• Position Encoder: An MLP processing the relative position vector to learn distance-attenuation relationships.
• Bounding Box Encoder: Encodes object locations and classes to model occlusion effects.

These features are fused and fed into two specialized prediction heads:

• PL Head: Takes fused image and position features to regress Path Loss.
• SH Head: Takes features from all three encoders to model Shadow Fading, capturing complex occlusion dynamics.

Training Strategy
The model is trained using a two-stage process to ensure stable convergence:

• Stage 1: The pre-trained MobileNet image encoder is frozen; only the remaining modules are trained.
• Stage 2: The entire framework is fine-tuned with a lower learning rate.
  The optimization uses a weighted sum of Mean Squared Errors (MSE) for both PL and SH components.

Key Results
Evaluated on the DeepSense 6G dataset (Scenarios 1–9), the proposed framework achieves state-of-the-art performance:

• Accuracy: The model reduces the Root Mean Squared Error (RMSE) for RSSI prediction by 50.3% compared to the previous baseline (reducing error from 1.41 dB to 0.70 dB under conventional conditions).
• Robustness: The model demonstrates exceptional stability against environmental interference. While the baseline's performance degrades significantly (RMSE increases from 0.78 dB to 1.41 dB) when distractors are present, the proposed model maintains an RMSE of 0.69 dB even with interference. Notably, the proposed model's performance with interference (0.70 dB) outperforms the baseline's best performance without interference (0.78 dB).
• Efficiency: The framework is remarkably lightweight, utilizing only 2.51 million parameters. This is approximately 19 times smaller than the recommended baseline model (47.7M parameters) and 34 times smaller than the baseline's highest-performing variant (86.1M parameters).
• Fine-Grained Impact: Ablation studies show that adding bounding boxes for small, semantically relevant objects (e.g., road signs) further improves prediction accuracy, reducing RSSI RMSE from 0.56 dB to 0.53 dB.

Significance and Claims
The paper claims that by aligning the learning objective with the underlying physics of signal propagation, the model significantly reduces learning difficulty and achieves superior interpretability. The authors assert that this approach establishes a new state-of-the-art by simultaneously delivering high accuracy, robustness to environmental noise, and computational efficiency. The combination of these factors makes the framework highly suitable for real-time, on-device deployment in edge scenarios, such as in-vehicle computing systems and unmanned aerial vehicles, paving the way for more intelligent and reliable wireless communication systems.