OAK

GIST Library Login

검색

Metadata Downloads

Abstract: Lane lines contain a lot of information beyond informing the direction of the vehicle's progress when driving on the road. The direction of driving of other vehicles and furthermore, traffic rules have meaning. So it is very important to detect it. In autonomous vehicles, lane detection is a task of providing essential information for vehicle control. Lane detection is used for vehicles to locate themselves on the road and maintain the correct path. However, lanes occupy a very small part of the overall image and have a wide variety of forms depending on the road environment. Detection is often difficult depending on weather conditions and the surrounding environment. In this paper, we propose a method to detect the overall lane by overcoming these challenges and restoring even the distorted part. This pipeline, which creates multiple lane candidates and allows each lane candidate to find the actual shape of the lane, consists of four modules: lane point regressor, lane point classifier, lane set suppressor, and lane set classifier. Through this pipeline, we introduce a novel methodology in which lane candidates correspond to all lanes in the image by selecting a label lane to correspond to by L1 distance calculation and having a shape that matches the label lane.

Table Of Contents: List of contents
Abstract i
List of contents ii
List of tables iv
List of figures v
I. INTRODUCTION 1
1. 1. Research background 1
1. 2. Previous research 1
1. 3. Proposed model 2
1. 4. Related works 3
1. 4. 1 Visual geometry approch 3
1. 4. 2. Lane Segmentation approch 3
1. 4. 3. Anchor based approch 3
II. DATASET 4
2. 1. TuSimple 4
III. Approch 5
3. 1. Model Pipleline 5
3. 1. 1. Lane representation 5
3. 1. 2. Points and set 6
3. 1. 3. Pipeline abstract 6
3. 2. Lane Point Regressor 8
3. 2. 1. Model 8
3. 2. 2. Label Lane Assignment 8
3. 2. 3. Training details 9
3. 3. Lane Point Classifier 11
3. 3. 1. Model 11
3. 3. 2. Classification of set points and labeling 11
3. 3. 3. Training details 12
3. 4. Lane Point Suppressor 13
3. 4. 1. Model 13
3. 4. 2. Overlapping lane candidate 14
3. 4. 3. Best lane selection 15
3. 5. Lane Set Classifier 15
3. 5. 1. False lane 15
3. 5. 2. False lane dataset 16
3. 5. 3. Model 17
3. 5. 4. Training details 18
IV. Experiment 19
4. 1. TuSimple 19
4. 1. 1. Data augmentation 19
4. 1. 2. Pipleline output 20
4. 1. 3. Effect of data augmentation 21
4. 1. 4. Inference 22
4. 1. 5. TuSimple Metric 24
V. Ablation Study 25
5. 1. Ablation for data augmentation 25
5. 2. Ablation for number of lane candidate 25
5. 3. Ablation for size of path 26
VI. Future Work & Conclusion 27
Summary 28
References 29

공개 및 라이선스

qrcode

OAK GIST Repository는 국립중앙도서관 OAK Repository 보급사업으로 구축되었습니다.