PaperlayerView exact revision on arXiv

Source-linked AI summary

LiDAR Aided Future Beam Prediction in Real-World Millimeter Wave V2I Communications

Shuaifeng Jiang, Gouranga Charan, Ahmed Alkhateeb

arXiv:2203.05548v1eess.SPcs.IT

TL;DR

MmWave and terahertz systems need narrow beams that make beam training costly, particularly for highly mobile links. This paper uses LiDAR sensory sequences with machine learning to predict current and future beams and evaluates the approach on DeepSense 6G vehicle-to-infrastructure data. The LiDAR-aided model achieves high beam accuracy while substantially reducing training overhead and performs comparably to a baseline with previous optimal-beam knowledge.

  • Problem

    Narrow-beam mmWave/THz communication requires costly beam training, motivating methods that reduce overhead in highly mobile scenarios.

  • Method

    A machine-learning model uses LiDAR sensing data to predict current and future beams in a vehicle-to-infrastructure setting, trained from sequential LiDAR and optimal-beam data.

  • Results

    95.6% top-5 accuracy for LiDAR beam prediction reduces beam training from 64 to 5, while achieving comparable performance to the baseline.

  • Takeaways & Limitations

    The results demonstrate the potential of LiDAR sensory information for real-world mmWave/THz beam prediction and tracking.

Abstract

from arXiv · show

This paper presents the first large-scale real-world evaluation for using LiDAR data to guide the mmWave beam prediction task. A machine learning (ML) model that leverages the LiDAR sensory data to predict the current and future beams was developed. Based on the large-scale real-world dataset, DeepSense 6G, this model was evaluated in a vehicle-to-infrastructure communication scenario with highly-mobile vehicles. The experimental results show that the developed LiDAR-aided beam prediction and tracking model can predict the optimal beam in $95\%$ of the cases and with more than $90\%$ reduction in the beam training overhead. The LiDAR-aided beam tracking achieves comparable accuracy performance to a baseline solution that has perfect knowledge of the previous optimal beams, without requiring any knowledge about the previous optimal beam information and without any need for beam calibration. This highlights a promising solution for the critical beam alignment challenges in mmWave and terahertz communication systems.

I. INTRODUCTION

The paper addresses costly beam training in mobile mmWave/THz systems by using LiDAR sensing and machine learning for current and future beam prediction. It evaluates this approach on large-scale real-world vehicle-to-infrastructure data, reporting high accuracy and substantially reduced training overhead.

  • Motivation: Narrow mmWave/THz beams require large antenna arrays and incur prohibitively large beam-training overhead, especially when mobility causes rapid channel changes.The motivation is to reduce or eliminate repeated training in highly mobile settings.
  • Motivation: LiDAR is investigated as an alternative sensing modality because wireless and radar sensing consume wireless resources, while camera sensing suffers under poor lighting and raises privacy concerns.Position-based sensing can also require extra signaling and create privacy concerns.
  • Contributions: The paper proposes LiDAR sensory data for both beam prediction and beam tracking, extending prior current-beam studies to future-beam prediction.The future-beam capability is identified as the paper’s first investigation of this use of LiDAR.
  • Contributions: An efficient machine learning model is proposed for the LiDAR-aided beam prediction tasks.The model is part of the paper’s contribution alongside the real-world evaluation.
  • Evaluation and contributions: 95% optimal-beam prediction with beam training overhead reduced from 64 to 5 demonstrates high-accuracy LiDAR-aided beam management.The evaluation uses a vehicle-to-infrastructure scenario from the DeepSense 6G dataset.

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system uses a LiDAR-equipped base station to sense a mobile user and environment, then formulates beam management around selecting beams from a predefined codebook. The received-signal model specifies the downlink observation used in this mmWave system.

  • II. SYSTEM MODEL AND PROBLEM FORMULATION: The section introduces LiDAR-aided current and future beam prediction as the two considered beam-management tasks.These tasks are the problem formulations built on the adopted system model.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: The base station senses the environment and moving user equipment with LiDAR, using the sensing information for beam prediction and tracking.Figure 1 also presents the proposed machine-learning model blocks.
  • A. System Model: The considered system is a mmWave downlink in which a base station serves a mobile user equipment.The base station is equipped with an antenna array and LiDAR sensor, while the user equipment is single-antenna.
  • A. System Model: The downlink received signal is modeled as y[t] = hH[t]f[t]s + n[t].Here h[t] is the channel vector, f[t] is the transmit beamforming vector, s is the transmitted complex symbol, and n[t] is receive noise.
  • A. System Model: The channel vector h[t] describes the BS–UE channel at time step t.It has dimension N×1, matching the base station’s N-element antenna array.
  • A. System Model: The transmit beamforming vector f[t] is selected from a predefined beamsteering codebook F = {f1, . . . , fM} of size M.The transmitted signal has power P, and the noise covariance is specified by σ².

B. Problem Formulation and ML Task Definition

The paper formulates LiDAR-aided current beam prediction and future beam tracking as ML tasks that select optimal beams from a predefined codebook. It also compares tracking with a baseline that uses previously selected beams.

  • The optimal beam is the codebook beam that maximizes receive power, and prediction can therefore target its index.
  • LiDAR beam prediction estimates the current optimal beam from LiDAR data available through the current time step.
  • LiDAR beam tracking estimates future optimal beams from LiDAR observations collected through the current time step.
  • The tracking lead-time v specifies how far into the future the beam is predicted, while current prediction is the special case v = 0.
  • The baseline predicts future beams from the previously used beam sequence, which contains information about UE movement patterns.

III. PROPOSED LIDAR AIDED BEAM MANAGEMENT SOLUTION

The proposed solution combines LiDAR sensing with machine learning for current and future beam management, and compares it with a beam-history baseline.

  • The approach uses LiDAR data for both beam prediction and beam tracking.
  • The paper describes a machine learning model for the LiDAR-aided prediction and tracking tasks.
  • A beam-history-based solution is included as the baseline for comparison.

A. Key Idea: Why LiDAR?

mmWave and THz systems require narrow beams and large antenna arrays, but beam adjustment creates substantial overhead that worsens in highly mobile settings. Because line-of-sight geometry helps determine the optimal beam, sensing can provide useful environmental awareness.

  • Large antenna arrays and narrow beams are needed to achieve sufficient receive power in mmWave/THz systems.
  • High beam-training overhead reduces spectral efficiency and increases latency, especially in highly mobile and dynamic scenarios.
  • The direction and position of communication terminals and surrounding geometry influence the optimal beam because transmission power is concentrated on line-of-sight paths.

B. Proposed LiDAR-Aided System Operation

At each time step, the base station captures LiDAR imagery and feeds a sequence of images to an ML model that predicts promising beams for serving the user.

  • The base station captures a LiDAR image at each time step, such as during a coherence time.
  • A sequence of L LiDAR sensing images is provided to an ML model.
  • The model predicts the top-k promising beams for serving the user.

C. Deep Learning Model

The model uses recurrent neural networks to extract sequential features from LiDAR and beam inputs, then classifies candidate beams with softmax scores. It is trained offline to predict current and future beam sequences.

  • Model architecture: An RNN architecture with embedding and GRU blocks extracts sequential features for current and future beam prediction.The model contains repeated embedding and gated recurrent unit components.
  • Model architecture: The embedding block compresses raw inputs and maps semantically similar inputs into nearby representations.The experimental embedding dimensions De and Me are both set to 64.
  • Feature extraction: The model uses W single-layer GRUs with zero-initialized hidden states and a hidden-state size of 64.The same RNN architecture is used across the three machine-learning beam-management tasks.
  • Classification: A softmax classifier outputs one score per codebook beam, with the highest-scoring beam selected as the prediction.The score vector contains M beam scores between 0 and 1.
  • System context: The system setup places a LiDAR sensor in front of the base-station transceiver to obtain sensing information.The setup includes a moving user equipment transmitter and mmWave base-station antenna arrays.
  • Training: The model is trained offline using supervised sequences of LiDAR or beam inputs and one-hot target beam outputs.The baseline uses beam-history inputs, and both models use cross-entropy loss on the last four outputs.

IV. EVALUATION SETUP

The evaluation uses DeepSense 6G, a large-scale real-world dataset containing co-existing multimodal sensing and communication data, to assess LiDAR-aided beam management.

  • Evaluation dataset: The evaluation is the first large-scale real-world assessment of LiDAR-aided beam prediction using the DeepSense 6G dataset.The dataset contains co-existing multimodal sensing and communications data.

A. DeepSense 6G Secnerio 8

DeepSense 6G Scenario 8 models a moving vehicle communicating with a fixed base station equipped with a 60 GHz phased array and LiDAR sensor.

  • Scenario setup: Scenario 8 consists of a moving user-equipment transmitter, a fixed base-station receiver, and a LiDAR sensor.The vehicle carries an omnidirectional 60 GHz transmitter, while the base station receives with a 60 GHz phased array.
  • Scenario setup: The base station uses a 16-element phased array with a predefined codebook of 64 beams.The LiDAR sensor is placed in front of the transceiver to sense the communication environment.

B. Devolopment Dataset Generation

The development dataset pairs LiDAR observations with optimal beam indices, uses leakage-free sequence-level splitting, and evaluates prediction accuracy and training overhead against a beam-history baseline.

  • Dataset construction: Each Scenario 8 sequence corresponds to one road pass, and optimal beam indices are extracted from received powers.The resulting samples pair LiDAR sensing data with optimal beam indices.
  • Dataset construction: The training format uses an observation window of W = 8 and predicts three future beams.Each sample contains eight LiDAR steps, eight corresponding optimal beams, and three future optimal beam indices.
  • Evaluation metric: Top-k accuracy measures the percentage of test samples whose ground-truth beam appears among the k highest-scoring predicted beams.This metric is used to evaluate the proposed and baseline approaches.
  • Prediction accuracy: 95.6% top-5 accuracy enables the LiDAR model to find the optimal beam while reducing beam training from 64 beams to 5.The baseline assumes perfect knowledge of the previous eight optimal beams, which requires high training overhead.
  • Beam tracking: The LiDAR approach maintains beam tracking from up-to-date sensing, whereas baseline accuracy degrades as the operation window lengthens.Without very frequent exhaustive training, the LiDAR-aided approach consistently outperforms the baseline.
  • Training overhead: With a three-step prediction window, the LiDAR-aided approach requires around 10% of the baseline training overhead.Figure 3 compares average one-step beam-training overhead across operation-window lengths using the top five beams.
  • Beam tracking: Figure 3 compares the two future-beam prediction approaches across different operation windows and highlights lower training overhead for the LiDAR-aided solution.The comparison concerns beam-tracking performance as the operation window changes.

VI. CONCLUSION

The LiDAR-aided machine-learning approach achieves high current- and future-beam prediction accuracy on DeepSense 6G while requiring only a small fraction of baseline beam-training overhead.

  • 95.6% top-5 accuracy is achieved for current-beam prediction on the DeepSense 6G real-world dataset.
  • 95.0% top-5 accuracy is achieved for first-future-beam tracking.
  • 10.4% beam-training overhead is sufficient for the LiDAR-aided approach to match the baseline performance.
  • The baseline has perfect knowledge of previous optimal beams, whereas the LiDAR-aided approach uses LiDAR data.
Loading 2203.05548v1…