A Study on Intelligent Vehicle-to-everything Communications for 6G = 6G 통신을 위한 지능형 차량-사물 간 통신에 관한 연구

다국어 초록 (Multilingual Abstract)

With the rapid advancement of intelligent transportation systems (ITS), vehicle-to-everything (V2X) communication has emerged as a key enabler for next-generation applications such as autonomous driving, cooperative perception, smart traffic control, and connected pedestrian safety.
The communication requirements for these V2X applications fundamentally differ from conventional wireless systems, which mainly focus on generic data transmission.
In contrast, V2X communication should satisfy stringent performance requirements such as ultra-low latency, high reliability, and massive device connectivity.
While 5G V2X has introduced several performance improvements, these are still insufficient to fully support the increasing demands of real-time decision-making, dense vehicular networks, and safety-critical scenarios.
As vehicular environments become more dynamic and complex in the 6G era, conventional V2X communications fail to meet these requirements.
This dissertation presents novel deep learning-based solutions to address the key challenges in 6G-enabled V2X communication systems, including ultra-low latency, high reliability, and massive connectivity.
To this end, we develop DL-driven frameworks that incorporate advanced technologies such as reconfigurable intelligent surfaces (RIS), multimodal sensing, and large multimodal models (LMMs).

In the first part, we propose a Transformer-based RIS phase shift control (T-RPSC) scheme designed to enhance RIS-assisted V2X communications. By leveraging the temporal correlations in RIS-aided channels, T-RPSC employs Transformer networks to predict future channel states and generate optimal RIS phase shifts, thereby ensuring reliable communication in dynamic vehicular environments.

In the second part, we propose a multimodal sensing-aided beam management (MMBM) framework for upper-mid and millimeter-wave (mmWave) wireless systems. MMBM utilizes data from LiDAR and RGB cameras to extract accurate 3D geometric information, facilitating efficient beamforming vector generation. This approach addresses the limitations of conventional codebook-based beam training, such as beam misalignment and latency from exhaustive search, by enabling robust and low-latency beamforming even under challenging conditions like reflective surfaces and occlusions.

In the third part, we present a quantized feature-enhanced LMM model for V2X communications (QF-LMMV2X). This novel framework integrates LMMs with quantized sensing feature feedback to enable adaptive V2X communication. Vehicles extract task-relevant features from multimodal sensing data and transmit the quantized features to roadside units (RSUs), which then utilize pre-trained LMMs to infer context-aware decisions for various vehicular tasks. This approach enhances the reliability and adaptability of V2X communication systems in complex and dynamic driving environments.

국문 초록 (Abstract)

지능형 교통 시스템(Intelligent Transportation Systems, ITS)의 급속한 발전과 함께, 차량-사물 간 통신(Vehicle-to-Everything, V2X)은 자율주행, 협력 인식, 스마트 교통 제어, 보행자 안전 등 차세대 애플리케이션의 핵심 기술로 주목받고 있다.
이러한 V2X 통신 시스템은 단순한 데이터 전송을 목표로 하는 기존 무선 시스템과 달리, 초저지연, 초고신뢰성, 대규모 연결성과 같은 엄격한 통신 성능 요구사항을 만족해야 한다.
5G V2X는 이러한 요구를 일정 부분 충족시켰지만, 실시간 의사결정, 고밀도 차량 네트워크, 안전 필수 환경을 완전히 지원하기에는 여전히 부족한 점이 존재한다.
특히 6G 시대에 접어들며 차량 환경이 더욱 동적이고 복잡해짐에 따라, 기존 V2X 통신 메커니즘만으로는 요구 조건을 만족하기 어렵다.
본 논문은 6G 기반 V2X 통신 시스템의 주요 과제들—초저지연, 고신뢰성, 대규모 연결성—을 해결하기 위한 딥러닝 기반 솔루션을 제시한다.
이를 위해 재구성 가능한 지능형 표면(reconfigurable intelligent surface; RIS), 멀티모달 센싱(multimodal sensing), 대형 멀티모달 모델(large multimodal models; LMMs) 등의 첨단 기술을 통합한 딥러닝 프레임워크를 개발하였다.

논문의 첫 번째 파트에서는, RIS 기반 V2X 통신의 성능을 향상시키기 위한 Transformer 기반 위상 제어 기법 (T-RPSC)을 제안한다.
본 기법은 RIS 반사 채널의 시간적 상관관계를 활용하여 Transformer 신경망을 통해 미래 채널 상태를 예측하고, 이에 기반해 최적의 위상 제어 값을 산출함으로써 동적 차량 환경에서도 신뢰성 높은 통신을 실현한다.

두 번째 파트에서는, 중상위 대역 및 밀리미터파 (mmWave) 무선 시스템을 위한 멀티모달 센싱 기반 빔 관리 (MMBM) 프레임워크를 제안한다.
본 기법은 LiDAR 및 RGB 카메라 센싱 정보를 기반으로 3차원 공간 정보를 추출하고, 이를 활용하여 효율적인 빔포밍 벡터를 생성함으로써 기존 코드북 기반 빔 트레이닝에서 발생하는 빔 오정렬 및 탐색 지연 문제를 해결한다.
특히, 반사체나 가림 현상이 존재하는 까다로운 환경에서도 안정적인 저지연 빔포밍을 지원할 수 있다.

세 번째 파트에서는, 차량이 추출한 멀티모달 센싱 특징을 양자화하여 RSU로 전송하고, RSU에서 사전학습된 LMM을 통해 차량 제어 명령을 생성하는 새로운 프레임워크인 QF-LMMV2X를 제안한다.
본 기법은 양자화 피드백과 대형 멀티모달 모델을 결합함으로써, 차량 환경의 복잡성과 동적 특성에도 적응 가능하고 신뢰도 높은 V2X 통신을 실현한다.

목차 (Table of Contents)

• Abstract i
• Contents iii
• List of Tables vi
• List of Figures vii
• 1 Introduction 1
• 1.1 Background 1
• 1.1.1 Basics of V2X Communications 2
• 1.1.2 Deep Learning-based V2X Communications 3
• 1.2 Contribution and Organization 4
• 2 Transformer-based RIS Phase Shift Control for Ultra-Low Latency V2X Systems 6
• 2.1 Introduction 7
• 2.2 RIS-aided V2X Systems 11
• 2.2.1 V2X System Model 12
• 2.2.2 RIS-aided V2X Channel Model 15
• 2.2.3 Channel Aging Effect in RIS-aided V2X Systems 17
• 2.2.4 RIS Phase Shift Control Problem Formulation 18
• 2.3 Transformer-based RIS Phase Shift Control 19
• 2.3.1 Basic of Transformer 21
• 2.3.2 Channel Prediction Network 24
• 2.3.3 Phase Shift Control Network 28
• 2.4 Practical Issues for T-RPSC Implementation 32
• 2.4.1 Synthetic Training Dataset Generation 32
• 2.4.2 Wireless Environment Compatible Training using Meta Learning 33
• 2.5 Simulation Results 35
• 2.5.1 Simulation Setup 35
• 2.5.2 Simulation Results 38
• 2.6 Summary 41
• 3 Vision and LiDAR-aided Beam Management for 6G mmWave Communications 43
• 3.1 Introduction 43
• 3.2 Multi-Modal Communication System . 49
• 3.2.1 Upper-mid/mmWave MISO System Model 49
• 3.2.2 Multi-modal Sensing Techniques: LiDAR and RGB camera 50
• 3.2.3 Beam Management in Conventional 5G NR 52
• 3.3 Multi-Modal Sensing-aided Beam Management 53
• 3.3.1 RoI Extraction 56
• 3.3.2 CV-based Geometric Channel Information Estimation 60
• 3.3.3 Beam Generation and Analysis of MMBM Beamforming Gain 63
• 3.3.4 Implementation Scenario of MMBM 65
• 3.4 Practical Issues for MMBM Implementation 66
• 3.4.1 LoS Path Blockage 66
• 3.4.2 Complexity Analysis 67
• 3.5 Simulation Results 68
• 3.5.1 Simulation Setup 68
• 3.5.2 Simulation Results 69
• 3.6 Summary 72
• 4 Large Multimodal Model-based V2X Communication with Quantized Feature
• Enhancement 74
• 4.1 Introduction 75
• 4.2 System Model of QF-LMMV2X Communications 79
• 4.2.1 Uplink V2X OFDM System Models 79
• 4.2.2 Basics of Large Multimodal Models 80
• 4.2.3 Quantization Techniques 82
• 4.3 Quantized Feature-enhanced Large Multimodal Model for V2X Communications 84
• 4.3.1 Sensing FE at VUE 85
• 4.3.2 LMM-based CG at RSU 90
• 4.3.3 Overall Procedure of QF-LMMV2X 92
• 4.4 Training Strategy for QF-LMMV2X 94
• 4.4.1 LoRA-Based Fine-Tuning of the LMM 95
• 4.4.2 Sequential Training of Sensing FE and LMM Head 96
• 4.4.3 Prompt Engineering in QF-LMMV2X 98
• 4.5 Numerical Results 100
• 4.5.1 Simulation Setup 100
• 4.5.2 Simulation Results 102
• 4.6 Summary 106
• 5 Conclusion 107
• A Proof of the Proposition 1 109
• Abstract (In Korean) 122
• 감사의 글 124