Lightweight Vision-Aided Beam Tracking for Cross-Environment mmWave Communications

Technical Summary: Lightweight Vision-Aided Beam Tracking for Cross-Environment mmWave Communications

Problem Statement
Millimeter-wave (mmWave) communications rely on highly directional transmission, necessitating accurate and low-overhead beam management. While sensing-aided beam tracking using contextual information (e.g., from cameras) has emerged as a promising solution to reduce beam training overhead, existing approaches face significant challenges in real-world deployment. Current methods often focus on predicting beams for a single time slot within a specific environment, neglecting model efficiency in terms of memory and computational complexity. Furthermore, most existing solutions lack robustness when deployed across diverse environments with varying geometries, traffic patterns, and background conditions. There is a critical gap in developing low-complexity, vision-aided frameworks capable of generalizing across different scenarios while performing long-term beam prediction (jointly predicting current and future beams).

Methodology
The paper proposes a lightweight, vision-aided framework formulated as a sequence-to-sequence (Seq2Seq) classification task. The system aims to jointly predict optimal beam indices for the current time slot and $J$ future time slots based on a sequence of past RGB visual observations.

• Architecture: The core of the proposed model is a lightweight neural network designed for efficient spatial-temporal feature learning.

  • Image Encoder: Instead of heavy backbones, the model utilizes a ResNet-style stem followed by four Depthwise Separable Convolution (DS-Conv) blocks. These blocks employ inverted residual structures (expanding channels via $1\times1$ convolutions, performing spatial downsampling via $3\times3$ depthwise convolutions, and projecting back via $1\times1$ convolutions) to reduce complexity while maintaining feature representation.
  • Temporal Modeling: The encoded features are processed by a Gated Recurrent Unit (GRU) and a Multihead Attention (MHA) module to capture temporal dependencies in the visual sequence.
  • Classifier: A three-layer Multilayer Perceptron (MLP) outputs beam selection probabilities for all candidate beams across the $J+1$ time slots.

• Enhanced Learning Strategies: To improve robustness and generalization in cross-environment settings, two specific strategies are introduced:

  1. Hierarchical Data Augmentation: This operates at both sequence and frame levels to emulate imperfect sensing conditions while preserving UE motion patterns.
     • Sequence-level: Morphological perturbations (erosion/dilation) and Gaussian blur are applied consistently across all frames in a sequence to simulate motion mask extraction errors and sensing uncertainty.
     • Frame-level: Additive Gaussian noise is introduced to handle local intensity fluctuations.
  2. Beam Power-Based Label Smoothing: Recognizing that multiple beams may offer comparable performance, the authors replace strict one-hot supervision with a softened distribution derived from beam power vectors. The total loss function combines a Focal loss (for hard labels) and a cross-entropy loss (for soft labels), weighted by a hyperparameter $\lambda$. This prevents the model from becoming over-confident on a single beam class.

Key Contributions

1. Cross-Environment Formulation: The work shifts the focus from single-environment beam tracking to a multi-scenario setting, formulating the problem as a Seq2Seq task to handle long-term prediction across geometrically distinct environments.
2. Lightweight Architecture: A compact model is developed using depthwise separable convolutions and efficient temporal modeling, significantly reducing computational burden compared to high-capacity baselines.
3. Robustness Mechanisms: The integration of hierarchical data augmentation and beam power-based label smoothing specifically targets the challenges of environmental variation and sensing imperfections.
4. Comprehensive Evaluation: The framework is evaluated on real-world data from two distinct DeepSense 6G scenarios (Scenario 9 and Scenario 31), which feature different Base Station locations, fields of view, and traffic patterns.

Experimental Results
The proposed model was tested on a merged dataset of DeepSense 6G Scenarios 9 and 31, predicting beams for the current slot and the next three ($J=3$) slots using a 64-beam codebook.

• Accuracy: The proposed model, when utilizing both data augmentation and label smoothing, achieved an average Top-1 accuracy of 38.82%, Top-3 of 73.01%, Top-5 of 84.03%, and an Average Distance-Based Accuracy (ADBA) of 83.97%.
• Comparison with Baselines:
  • The proposed model outperformed the state-of-the-art CNN-GRU solution (specifically designed for a single scenario) in all metrics under the best settings.
  • While a high-capacity ResNet-GRU baseline achieved slightly higher absolute accuracy (e.g., 41.69% Top-1), it did so at a significantly higher computational cost.
• Complexity Reduction: Compared to the ResNet-GRU baseline, the proposed lightweight model reduced the number of parameters by a factor of approximately 52 (from 11.3M to 0.217M) and computational complexity (FLOPs) by a factor of approximately 79 (from 12.2G to 155M).
• Temporal Performance: The model maintained robust performance across all four time slots ($t$ to $t+3$), with Top-5 accuracy remaining above 83.5% and DBA scores around 84%, demonstrating its capability for long-term prediction despite the reduced model size.

Significance
The paper claims that its primary significance lies in demonstrating that lightweight sensing-aided models, when equipped with appropriate regularized learning strategies (hierarchical augmentation and label smoothing), can maintain satisfactory beam tracking performance across diverse environments. The work addresses the trade-off between prediction accuracy and deployment feasibility, showing that substantial reductions in model size and computational complexity can be achieved without sacrificing the robustness required for cross-environment mmWave communications. The results suggest that such efficient models are viable for practical deployment where computational resources are constrained, offering a scalable alternative to heavy, environment-specific deep learning solutions.