|
Title |
3차원 GPR을 이용한 도로 하부 공동 자동탐지 체계의 개발과 현장 실증: 재현율을 유지 하는 YOLO 기반 검토량 저감
|
|
Authors |
허춘근(Choun-Kun Hur) ; 박정찬(Jung-Chan Park) ; 이창훈(ChangHoon Lee) ; 송수민(Su-Min Song) ; 오홍섭(HongSeob Oh) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.5.128 |
|
Keywords |
딥러닝; YOLO; 지표투과레이더(GPR); 하이퍼볼라; 지하 공동; 현장 실증 Deep learning; YOLO; Ground-penetrating radar (GPR); Hyperbola; Underground cavity; Field validation |
|
Abstract |
In Korea’s underground safety management, roads are screened for cavities with vehicle-mounted ground-penetrating radar (GPR), and suspected locations are confirmed by drilling and endoscopy. Deep learning is now widely used in that screening step. In the vertical radar section (B-scan) a cavity appears as a hyperbola, and so does a buried pipe, so a detector that finds hyperbolas reliably still leaves the analyst many non-cavity candidates.
Raising the threshold would cut that work but lose cavities, so this study instead removes only candidates that are unlikely to be cavities. YOLOv7x detects the hyperbolas. For each detected box, a second YOLOv7x model examines the horizontal C-scan slices across the 20 channels over the box depth and counts the slices showing a cavity or a pipe. A gradient boosting model combines these counts with the detection confidence and with features computed from the raw traces, such as polarity relative to the surface reflection, reflection strength and cross-channel continuity, and low-scoring locations are dropped from the review list. On their own, the C-scan cavity detections and the detection confidence separated confirmed cavities best, whereas polarity and reflection strength were weak (AUC 0.58?0.68). Evaluation used Korea Authority of Land & Infrastructure Safety surveys: 441 lines, 467.0 km, 344 sites, 2019?2025, 323 drilling-confirmed cavities, with training and test sites kept separate. At a detection confidence of 0.30, the detector found 304 of 323 confirmed cavities (94.1%) at 294.9 boxes per km. Merging repeated boxes over one spot gives 80.7 review locations per km, which the score lowered to an estimated 65.6, 18.7% less, with all 304 cavities kept in this internal evaluation.
|