Efficient Trajectory Prediction via Interaction Pruning in Autonomous Driving Scenes. Muhammad Atta ur Rahman Autonomous Driving Intelligence Research Section (ADIR), ETRI = 자율 주행 장면에서 상호작용 가지치기를 통한 효율적인 궤적 예측

국문 초록 (Abstract)

상호작용가지치기를통한자율주행환경에서의효율적인궤적예측정확하 고효율적인궤적예측은복잡하고역동적인교통상황에서안전하게주행하 기위한자율주행시스템의핵심역량중하나이다.본학위논문은자율주행 을위한모션예측모델의상황인식능력과계산효율성을향상시키는일반 적인 프레임워크를 제안한다. 논문의 첫 번째 파트에서는 궤적 예측에 차선 중심선만사용하는기존접근방식을넘어,차선경계및도로가장자리와같 은고밀도벡터맵정보를통합하는새로운접근법인 LANet을제안한다.이 디자인은도로장면에대한더욱통합된공간인식을가능하게하며,향상된 맵표현으로인한계산량증가문제를해결하기위해 CAIP라는관련성기반 맵 가지치기 메커니즘을 도입한다. CAIP는 대상 에이전트와 가장 관련 있 는 맵 세그먼트만을 필터링함으로써 효율성을 높인다. 두 번째 파트에서는 DPG-Traj라는 그래프 기반 예측 모델을 제안한다. 이 모델은 방향 기반 근 접 그래프 가지치기(Directional Proximity Graph Pruning)를 통해 에이전트 간 및 에이전트-지도 간 상호작용에서 불필요한 연결을 동적으로 제거하여 모델 복잡도를 줄인다. 이를 통해 모델은 가장 유의미한 동적 및 의미론적 관계에집중할수있게된다.예측정밀도는향상되고계산비용은절감되어, DPG-Traj는 실시간 응용에 더욱 적합하다. Argoverse 2 모션 예측 데이터셋 을활용한대규모실험을통해두접근법의효과성을검증하였다.본논문의 기여는실시간성과맥락인식을모두고려한확장가능한프레임워크를수립 함으로써자율주행시스템의궤적예측분야를한단계진보시킨다. 키워드:자율주행, LANet, DPG-Traj,연결가지치기, Argoverse 2.

다국어 초록 (Multilingual Abstract)

Accurate and efficient trajectory prediction is one of the fundamental capabilities of autonomous driving systems for navigating safely through difficult and dynamic traffic scenarios. This thesis formulates a general framework enhancing the contextual awareness and computational efficiency of motion prediction models. In the first part of this thesis, we present LANet, a novel approach incorporating dense vector map features, such as lane boundaries and road edges, into trajectory prediction. This design builds upon the conventional reliance on lane centerlines and facilitates more integrated spatial understanding of the road scene. To manage the increased computational overhead of representing more map elements, we propose CAIP, a relevance-based pruning mechanism that filters only the most pertinent map segments to the target agent. In the second part, we present DPG-Traj, a graph-based prediction model that alleviates model complexity through Directional Proximity Graph Pruning. It removes unnecessary connections in agent-to-agent and agent-to-map interactions dynamically so that the model can focus on the most informative dynamic and semantic relationships. With improved forecasting precision and less computation expense, DPG-Traj is better suited for real-time applications. Largescale experiments on the Argoverse 2 motion forecasting dataset validate the effectiveness of both approaches. The combined contributions of this thesis advance the state of the art in motion prediction a step further by establishing a scalable and context-aware framework for real-world autonomous driving systems.

Index Terms: Autonomous Driving, LANet, DPG-Traj, Connection pruning, Argoverse 2

목차 (Table of Contents)

• List of Figures v
• List of Tables xi
• 1. INTRODUCTION 1
• 1.1 Background and Motivation 1
• 1.2 Limitations of Existing Models 2
• 1.3 Research Objectives and Thesis Contributions 3
• 1.3.1 LANet: A Lane Boundaries-Aware Approach for
• Robust Trajectory Prediction 3
• 1.3.2 DPG-Traj: Enhancing Trajectory Prediction with
• Directional Prox- imity Graph Pruning 4
• 1.4 Experimental Evaluation 4
• 1.5 Organization of the Thesis 5
• 2. Literature Review 7
• 2.1 Symmetric Scene Representation 7
• 2.2 Encoding of Context and Dynamics 8
• 2.2.1 Encoding Agent Dynamics over Time 8
• 2.2.2 Map Feature Representation 9
• 2.2.3 Transformer-based Encoding 9
• i
• 2.2.4 Agent-to-Agent Interaction 10
• 2.2.4.1 Spatio-Temporal Graph Neural Networks 11
• 2.2.4.2 Self-Supervised Learning Methods 11
• 2.2.4.3 Causal and Influence-Aware Interaction
• Modeling 12
• 2.2.4.4 Transformer-based Architectures 13
• 2.2.5 Agent-to-Map Interaction 14
• 2.2.5.1 Graph Neural Networks (GNNs) 15
• 2.2.5.2 Cross-attention mechanisms 16
• 2.2.5.3 Joint Map and Agent Encoders Based on
• Transformers 17
• 2.2.6 Interaction Pruning Strategies in A2A and A2M . 18
• 2.2.6.1 Structural Pruning 18
• 2.2.6.2 Context-Aware Pruning 19
• 2.2.6.3 Attention-Based Pruning 20
• 3. Methodology 24
• 3.0.1 Problem Formulation 24
• 3.0.2 Overview of QCNet 25
• 3.0.3 Limitations of the QCNet and existing trajectory
• prediction methods 27
• ii
• 4. LANet 30
• 4.0.1 Methodology 30
• 4.0.1.1 Map Encoding 30
• 4.0.1.2 Agent Features Encoding 32
• 4.0.2 Context-Aware Interaction Pruning 35
• 4.0.3 The Decoder 39
• 4.0.4 Training Objectives 42
• 5. DPG-Traj 43
• 5.1 Methodology 43
• 5.1.1 How is it changed from LANet? 43
• 5.1.1.1 Map Representation 44
• 5.1.1.2 Agent Encoder Upgrades 45
• 5.1.1.3 Decoder Framework 46
• 5.1.2 Agent-to-Map Interaction Modeling with DPG-A2M 46
• 5.2 Agent-to-Agent Interaction Modeling with DPG-A2A 51
• 6. Experimental Results 57
• 6.0.1 Dataset 57
• 6.0.2 Evaluation Metrics 60
• 6.0.3 LANet Experimental Results 61
• 6.0.3.1 Quantitative Results Comparison 61
• iii
• 6.0.3.2 Ablation Study 62
• 6.0.3.3 Qualitative Results 63
• 6.0.4 DPG-Traj Experimental Results 65
• 6.0.4.1 Quantitative Results 65
• 6.0.4.2 Ablation Study 66
• 6.0.5 Qualitative Results 72
• 6.1 Conclusion 74
• Bibliography 76
• iv