EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

From Cracks to Craters: A Computer Vision Pipeline for Forecasting Pavement Degradation

Derek Hu, William Liu

American Journal of Capstone Research · 2026

DOI: 10.70251/1432 ↗

Kurz gefasst

Fachartikel in der Zeitschrift American Journal of Capstone Research, 2026. Verfasst von Derek Hu und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Potholes and related pavement distresses pose significant safety and financial consequences, yet municipal road maintenance remains largely reactive. Most computer vision frameworks operate exclusively on static imagery and do not project future degradation. To bridge this gap, this paper presents a two-phase pipeline combining automated visual detection with image-space degradation forecasting. A YOLOv8m model was trained on the PaveTrack_OD China subset, comprised of six distress types (transverse cracks, patched cracks, potholes, patched potholes, alligator cracks, and patched alligator crack) and four road features (manholes, street waste, speed bumps, puddles), achieving an overall mAP@0.50 of 0.461 on validation imagery. In the second phase, a Random Forest Regressor was trained on longitudinal bounding-box data from PaveTrack_PD to forecast future 2D pothole bounding-box dimensions (width and height in pixels) from initial dimensions and elapsed time over a six-month monitoring window. On controlled holdout test data, the regressor achieved values of 0.628 for width (MAE: 86.68 pixels) and 0.448 for height (MAE: 23.69 pixels), with initial bounding-box width identified as the dominant predictive feature (53.4% aggregate Mean Decrease in Impurity). To evaluate cross-domain generalizability, the pipeline was applied to 62 uncalibrated Google Street View image pairs from Seattle and Southern California. Evaluated at the image-pair level, YOLOv8m successfully detected and bounded the co-located target distress across both temporal captures in 53 of the 62 image pairs. However, downstream regression accuracy degraded under uncalibrated conditions, yielding a Median Absolute Percentage Error of 78.6% for width and 47.9% for height. These errors reveal fundamental constraints stemming from uncalibrated optical perspective shifts, inconsistent bounding-box aggregation, and multi-year extrapolation beyond the six-month training boundary. Resolving these bottlenecks via fixed-perspective cameras and spatial scale normalization is necessary to transition visual defect forecasting into proactive municipal infrastructure management.

Zitieren

APA

Hu, D., & Liu, W. (2026). From Cracks to Craters: A Computer Vision Pipeline for Forecasting Pavement Degradation. American Journal of Capstone Research. https://doi.org/10.70251/1432

Vancouver

Hu D, Liu W. From Cracks to Craters: A Computer Vision Pipeline for Forecasting Pavement Degradation. American Journal of Capstone Research. 2026. doi:10.70251/1432

BibTeX
@article{Hu2026,
  author = {Hu, Derek and Liu, William},
  title = {From Cracks to Craters: A Computer Vision Pipeline for Forecasting Pavement Degradation},
  journal = {American Journal of Capstone Research},
  year = {2026},
  doi = {10.70251/1432}
}
Kostenlos starten →