Sewer pipelines play a vital role in maintaining urban sanitation and public health, contributing to sustainable development. Early detection and maintenance of physical defects can reduce management costs and improve performance. Traditional CCTV inspections, consisting of two main stages: on-site video collection and office evaluation, face limitations such as technician fatigue and time consumption. To address these issues, a YOLO-based deep learning model for automated defect detection has been proposed. Using sewer pipeline interior images provided by AI Hub, this study classified eight defect types, including joint offsets, connector protrusions, and sediment deposits. The model achieved over 90% mAP accuracy (based on an IOU of 0.5) regardless of lighting and background noise. This technology is expected to enhance the efficiency of sewer maintenance, contributing to safer and more sustainable urban environments.
νμλ μμ€ν μ λμν λ° μΈκ΅¬ λ°μ§ μ§μμ μ§μ κ°λ₯ν λ°μ μ μμ΄ μ€λν μν μ μνν΄μλ€. μ΄μ κ°μ νμλ μμ€ν μ μ€μμ±μ μΈμ§νκ³ , μ§λ°© μ λΆ λ° μμΉ λ¨μ²΄λ νμ μμ€ν μ ν¨μ¨μ κ΄λ¦¬μ μ μ§λ₯Ό μνμ¬ 2024λ μ 2μ‘° 7,692μ΅ μμ μμ°μ μ± μ νμκ³ 2023λ λλΉ 5,567μ΅ μ(25%) μ¦κ°νμλ€(νκ²½λΆ, 2024). νμλ μμ€ν μ μ§λ¨κ³Ό μ μ§ λ³΄μλ ꡬ쑰물μ λ¬Έμ λ₯Ό μ μν λ°κ²¬Β·ν΄κ²°νμ¬ μ±λ₯μ κ°μ νκ³ κ΄λ¦¬ λΉμ©μ μ κ°νλ©°, μμ μ±κ³Ό μ΅μ μ μλΉμ€ μ 곡μ λͺ©μ μΌλ‘ νλ€.
νμ¬ CCTV κ²μ¬λ νμλ μμ€ν μ μ§λ¨ λ° μ μ§ λ³΄μ κ³Όμ μμ νμμ μΈ λκ΅¬λ‘ μΈμλκ³ μλ€. λ μ΄μ κΈ°λ° μμ€ν , μ΄μν μΌμ, μ μΈμ μ΄μ μΉ΄λ©λΌ λ± λ€λ₯Έ κΈ°μ μ μ κ·Όλ²κ³Ό λΉκ΅νμμ λ, CCTV κ²μ¬λ λΆμμ μ©μ΄ν μκ°μ μλ£λ₯Ό μ 곡νλ€λ μ₯μ μ κ°λλ€. CCTV κ²μ¬ κ³Όμ μ λΉλμ€ μμ§κ³Ό μ¬λ¬΄μ€μμ νλ ¨λ κΈ°μ μμ μν λΉλμ€ λΆμμΌλ‘, λ μ£Όμ λ¨κ³λ‘ ꡬμ±λλ€.
2022λ κΈ°μ€ μ΄κΈΈμ΄ 16λ§ 8,786kmμΈ νμκ΄λ‘μ κ²°ν¨ λ°κ²¬μ μν΄ μμλλ μκ°κ³Ό κΈ°μ μμ νΌλ‘λ νκ°μ ν¨μ¨μ±μ λΆμ μ μΈ μν₯μ λ―ΈμΉ μ μλ€.(νκ²½λΆ, 2024). μ΅κ·Ό λ₯λ¬λ κΈ°μ μ μ μ©μ΄ νλλ¨μ λ°λΌ, μ΄λ¬ν λ¬Έμ λ₯Ό ν΄κ²°νκΈ° μν΄ λ§μ μ°κ΅¬μλ€μ΄ μ»΄ν¨ν° λΉμ κΈ°μ μ μ΄μ©ν μλν μ κ·Ό λ°©μμ λͺ¨μνκ³ μλ€(μμν et al., 2018), (Cheng et al., 2018). νμ§λ§ μ ν μ°κ΅¬μ κ²½μ°λ Two stage detectorλ‘ μΆλ‘ μλκ° λλ € μ€μκ° νμ§κ° λΆκ°λ₯νλ€λ νκ³κ° μ‘΄μ¬νλ€.
νΉν, βYou Only Look Once (YOLO)β μκ³ λ¦¬μ¦μ κ°μ²΄ κ°μ§ κΈ°μ μ€ νλλ‘, κ³ μ μ²λ¦¬μ λμ μ νλλ₯Ό κ²ΈλΉν΄ μ€μκ° νμκ΄λ‘ κ²°ν¨ νμ§μ νΉν μ ν©νλ€(Redmon, J. et al., 2016). λ°λΌμ, λ³Έ λ Όλ¬Έμμλ YOLOv8μ νμ©νμ¬ νμκ΄λ‘μ κ²°ν¨μ νμ§νλ μμ€ν μ μ μνλ€. λ³Έ μμ€ν μ μ΄λ―Έμ§λ₯Ό λ€μ΄μ¬μ΄μ§νμ¬ κΈ°μ‘΄λ³΄λ€ νμ΅ λ° μΆλ‘ μλλ₯Ό μ¦κ°μν€λ©°, μ€μκ° νμ§λ₯Ό ν΅ν΄ κΈ°μ‘΄μ μλ κ²μ¬ λ°©μμ νκ³λ₯Ό 극볡νκ³ , νμλ κ΄λ¦¬μ ν¨μ¨μ±κ³Ό μ νμ±μ κ°μ νλ νμ μ μΈ μ κ·Ό λ°©μμ μ μνλ€.
μ 체 μ΄λ―Έμ§ λ€μ΄λ‘λ ν΄λ¦
Sewerpipe
βββ Dataset
βββ train
βββ CC
βββ CL
βββ ...
βββ ...
βββ val
βββ test
βββ Dataset(+Background)
βββ train
βββ ...
βββ test
mAP50 | Recall | Inference Time |
---|---|---|
0.7474 | 0.7001 | 11.1ms |
Optimal Parameters
Batch size | Optimizer | Cos_lr | Pretrained | Background | Label Smoothing |
---|---|---|---|---|---|
128 | Adam | False | True | X | 0.3 |
@inproceedings{title={Development of a real-time sewer pipe defect detection algorithm using YOLO},
author={Jo, Yeongjoo and Yeo, Sihyeong and Kim, Changwoo},
booktitle={2024 곡λνμ λ°νν}
year={2024},
month={March},
organization={Korean Society on Water Environment (νκ΅λ¬Όνκ²½νν), Korean Society of Water & Wastewater (λνμνμλνν)},
}
This project is licensed under the Apache License 2.0. See the LICENSE file for details.