How to Make a Braking Strategy about Invisible Obstacles : Autonomous Driving in Blind Spot
- 주제(키워드) Autonomous Driving , Blind Spot , Braking Strategy
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Young-June Choi
- 발행년도 2025
- 학위수여년월 2025. 8
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000034905
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
A blind spot is an area around a vehicle that the driver can not see directly or through mirrors, posing challenges in detecting vehicles, pedestrians, and obstacles. This visibility issue is more pronounced on curved roads due to visual obstructions and the angle of the road surface. Addressing blind spots in autonomous driving environments on curved roads is crucial for developing effective braking strategies. This paper presents a novel approach that integrates Imitation Learning(IL), Dataset Aggregation(DAgger) and Deep Reinforcement Learning(DRL) to formulate a braking strategy for invisible obstacles. Initially, the method uses expert dataset to train basic autonomous driving and safety braking strategies through IL. DAgger then refines the dataset by aggregating action outcomes and filtering out ineffective policies. Lastly, the fine-tuning process through DRL enhances learning speed using the validated dataset. Utilizing the Deep Deterministic Policy Gradient(DDPG) algorithm, the proposed model measures safety driving speed and distance, executing an optimal vehicle braking strategy. This comprehensive approach results in a significant reduction in collision rate by 26.2% on curved roads.
more목차
1. Introduction 1
1.1 Blind Spot 1
1.2 Autonomous Vehicle System 2
2. Related Work 4
2.1 Path Planning 4
2.2 Safety Distance Configuration 5
2.3 Reward Function Configuration 5
2.4 Autonomous Driving Model 6
3. Methodology 7
3.1 Dataset 7
3.2 Imitation Learning 9
3.3 DAgger 13
3.4 Deep Reinforcement Learning 15
4. Proposed Model 17
4.1 System Model Architecture 17
4.2 Proposed Braking Strategy 18
4.2.1 Braking Strategy 19
4.2.2 Reward Function 20
4.2.3 Algorithm for making a Braking Strategy 21
5. Simulation Results 23
5.1 Learning Efficiency and Performance 23
5.2 Braking Events 25
5.3 Collision Rate 27
6. Conclusion 29
7. Reference 30
