Title

Study on Degradation of Autonomous Driving Performance in Adverse Weather Conditions and Deep Learning-based Improvement

Alternative Title

강우 환경에서의 자율주행 인지 및 주행 성능 저하 분석과 딥러닝 기반 개선 연구

DGIST Authors

Hyeonjae Jeon ; Yongseob Lim

Advisor

임용섭

Issued Date

2026

Awarded Date

2026-08-01

Type

Thesis

Description

Autonomous driving, Rainy condition, Sensor blockage, Rainy image generation, Image restoration

Table Of Contents

Ⅰ. Introduction 1
1.1 Background and Motivation 1
1.2 Research Scope and Problem Definition 4
1.3 Research Objectives 6
1.3.1 Quantitative Analysis of Rain-induced Driving Behavior Degradation 6
1.3.2 Controllable Rainy Image Generation for Autonomous Driving Perception 7
1.3.3 Rain-aware Restoration for Perception and Driving Performance Evaluation 8
1.4 Dissertation Framework 9
1.5 Contributions 11
1.6 Organization of the Dissertation 14
1.7 Acronyms and Terminology 15
Ⅱ. Related Works 16
2.1 Autonomous Driving Perception under Adverse Weather Conditions 16
2.2 Sensor Blockage and Rain-induced Degradation 18
2.3 Simulation-based Evaluation for Autonomous Driving 19
2.4 Rainy Image Generation and Image-to-Image Translation 20
2.5 Multi-task Perception for Autonomous Driving 22
2.6 Image Restoration and Deraining 23
2.7 Diffusion-based Image Restoration and Degradation-aware Conditioning 25
2.8 Spatial Severity Modeling and Distortion–Perception Trade-off 26
2.9 Summary and Research Gap 28
Ⅲ. Driving Behavior Degradation Analysis under Rain-induced Sensor Blockage 29
3.1 Overview 29
3.2 CARLA-ROS-based Evaluation Framework 30
3.3 Rain-induced Sensor Blockage Modeling 33
3.4 Establishing the Rainfall Severity Equation 35
3.5 Lane Detection Algorithm 38
3.6 Steering Angle Estimation using Pure Pursuit 42
3.7 Experimental Setup 44
3.8 Evaluation Metrics 46
3.9 Results and Analysis: Straight Road Scenario 47
3.10 Results and Analysis: Curved Road Scenario 50
3.11 Discussion 53
3.12 Limitations 55
3.13 Chapter Summary 57
Ⅳ. Controllable Rainy Image Generation using Rain Style Diversification 58
4.1 Overview 58
4.2 Motivation for Controllable Rainy Image Generation 59
4.3 Overall RainSD-based Image Generation Framework 60
4.4 Rain Style Diversification Module 61
4.5 Dataset Construction using BDD100K 63
4.6 Qualitative Evaluation of Rainy Image Generation 64
4.7 Multi-task Perception Evaluation 67
4.8 Quantitative Results and Analysis 69
4.9 Discussion 70
4.10 Limitations 72
4.11 Chapter Summary 73
Ⅴ. Severity-Aware Diffusion Restoration for Robust Autonomous Driving Perception 74
5.1 Overview 74
5.2 Motivation for Severity-Aware Diffusion Restoration 76
5.3 Preliminary Study: FiLM-based Degradation Conditioning 78
5.4 Degradation Prior Extraction using DA-CLIP 82
5.5 Cross-Attention Conditioning for Rain Degradation Restoration 83
5.6 Spatial Severity Map Modeling 87
5.7 Hybrid Severity Map for Restoration Behavior Control 92
5.8 Distortion–Perception Trade-off in Rain Restoration 95
5.9 Time-aware Severity Scheduling 96
5.10 Proposed Severity-Aware Diffusion Restoration Framework 97
5.11 Experimental Setup 98
5.12 Driving-oriented Evaluation Plan 99
5.13 Discussion 100
5.14 Chapter Summary 102
Ⅵ. Integrated Evaluation and Discussion 103
6.1 Overview 103
6.2 Unified Rain-centered Research Framework 104
6.3 Connection between Image-level Degradation and Driving Behavior 106
6.4 Connection between Rainy Image Generation and Perception Evaluation 107
6.5 Connection between Restoration and Autonomous Driving Perception 108
6.6 Practical Implications 109
6.7 Limitations 110
6.8 Chapter Summary 111
Ⅶ. Conclusion and Future Works 112
7.1 Conclusion 112
7.2 Summary of Contributions 113
7.3 Future Works 115
7.4 Final Remarks 116
References 118
Abstract in Korean 121

URI

https://scholar.dgist.ac.kr/handle/20.500.11750/60739
http://dgist.dcollection.net/common/orgView/200001018438

DOI

10.22677/THESIS.200001018438

Degree

Doctor

Department

Department of Robotics and Mechatronics Engineering

Publisher

DGIST

Show Full Item Record