<link rel="stylesheet" href="styles.cf390c6bfbe7555f.css">
dc.contributor.authorKalluri, Deekshitha Chowdary
dc.date.accessioned2026-11-19T16:52:30Z
dc.date.graduationmonthDecember
dc.date.issued2026
dc.description.abstractState transportation agencies increasingly collect roadway imagery at a scale that makes frame-by-frame manual annotation expensive to sustain. The primary contribution of this thesis is a reproducible pavement-distress auto-labeling and retraining pipeline for agency-style roadway imagery. The system combines YOLO11x detection, ByteTrack temporal association, temporal-persistence filtering, representative-frame selection, GPS-overlay OCR, quality assurance, and retraining. A controlled experiment evaluates whether the candidate-selection stages can reduce roadway training-data volume without materially changing downstream detector performance. Nine pavement-distress sources were harmonized into a six-class corpus containing 98,661 retained images and 182,308 bounding-box annotations. The final experiment used a fixed public base of 78,928 training images, 9,866 validation images, and 9,867 frozen test images. Four YOLO11x conditions were compared under the same training and evaluation protocol: public-only control; filtered/subsampled training with 1,060 KDOT auto-labeled frames; equal-size random sampling of 1,060 frames from the corrected detector-positive candidate pool; and corrected unfiltered training with 3,266 KDOT detector-positive frames. On the frozen public test set, the control, filtered, equal-size random, and unfiltered conditions achieved mAP50 values of 0.665, 0.664, 0.661, and 0.667 and mAP50–95 values of 0.383, 0.381, 0.381, and 0.383, respectively. The filtered condition exceeded the same-size random condition by 0.003 mAP50, with equal mAP50–95 to three decimals. The filtered/subsampled condition therefore used 67.5% fewer added KDOT training frames than the unfiltered condition while differing by only 0.003 mAP50 and 0.002 mAP50–95. Because temporal-persistence filtering and representative subsampling changed together, the experiment evaluates the combined candidate-selection strategy rather than the isolated causal effect of temporal persistence. A single training seed was used, so the small metric differences are not presented as statistically significant. Qualitative review showed both the value and the limitation of temporal selection: an isolated false crack response was absent from the final selected arm, while a false Patch prediction on an adjacent vehicle survived selection. The result is therefore operational rather than an accuracy-improvement claim: the implemented workflow substantially reduced added roadway training-data volume while preserving essentially comparable performance on the common public-domain test set, and retained pseudo-labels still required quality assurance.
dc.description.advisorLior Shamir
dc.description.degreeMaster of Science
dc.description.departmentDepartment of Computer Science
dc.description.levelMasters
dc.identifier.urihttps://hdl.handle.net/2097/47594
dc.language.isoen_US
dc.subjectPavement distress detection
dc.subjectComputer vision
dc.subjectDeep learning
dc.subjectAutomated annotation
dc.subjectRoad infrastructure monitoring
dc.subjectTransportation infrastructure
dc.titleDevelopment and evaluation of a vision-based pavement distress detection and auto-labeling framework for transportation agency pavement monitoring
dc.typeThesis

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
DeekshithaChowdaryKalluri2026.pdf
Size:
15.28 MB
Format:
Adobe Portable Document Format

License bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
license.txt
Size:
1.65 KB
Format:
Item-specific license agreed upon to submission
Description: