Video-based Tire Wear Estimation System and Its Serving Method
- 주제(키워드) Tire Wear Estimation , Image Segmentation
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Hyung Il Koo
- 발행년도 2025
- 학위수여년월 2025. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000034472
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Tire wear is a crucial factor influencing vehicle performance and passenger safety, necessitating effective monitoring solutions. While traditional methods for measuring groove depth offer high accuracy, they are impractical for everyday drivers. Also, existing sensor-based and intelligent tire solutions often require specialized equipment, limiting their accessibility. This paper introduces a novel Tire Wear Estimation (TWE) system uses mobile phone cameras to estimate individual groove depths through close-up tire videos. Given the absence of suitable datasets capturing detailed tire groove information, we collect a large number of tire videos and manually measured groove depths to develop and evaluate our system. For semantic segmentation, we use a U-Net-based model, selected based on its balance between performance and computational efficiency, and apply a post-processing to refine the output masks. Using frame-wise tire masks extracted from input videos, we measure the width and depth of each dent—indented regions from the masks. By tracking these dimensions across frames, we estimate the actual groove depths. To ensure the system performance, we implement a model lifecycle-driven approach, allowing iterative updates based on acquisition of videos. Additionally, we develop an partial-automated mask quality verification method for efficient data validation, addressing the challenge of manual inspection for large-scale usage. Our proposed TWE system demonstrates an absolute error of 0.89 mm on generalization test videos, with an average latency of 2.44 seconds for 10-second tire videos, making it an efficient and practical real-world applications.
more목차
1 Introduction 1
2 Related Work 5
2.1 Semantic segmentation 5
2.2 Serving of Deep learning model 7
3 Proposed Method 9
3.1 Notation and overview of proposed method 10
3.2 Frame wise tire segmentation 12
3.2.1 Purpose of semantic segmentation 13
3.2.2 Segmentation mask post-processing 15
3.2.3 Segmentation mask dent localization by frames 16
3.3 Segmentation mask partitioning 21
3.3.1 Keyframe detection 21
3.4 Groove depth estimation 24
3.4.1 Interior-block width estimation 24
3.4.2 Exterior-block width estimation 25
3.4.3 Groove depth and width estimation 26
3.5 Model-lifecycle driven serving 27
4 Experiments and Results 29
4.1 Semantic Segmentation performance 29
4.1.1 Datasets 30
4.1.2 Implementation Details 32
4.1.3 Comparison results of Segmentation models 33
4.2 Video pre-screening 35
4.3 TWE system Serving result 36
4.3.1 1st Phase Update 37
4.3.2 Second Update 38
4.4 Uneven wear detection 38
4.5 Computation time 40
5 Conclusions 43
