STRUCTURACONNECTED DIGITAL ENGINEERING
Sign inStart free
STRUCTURA ACADEMIC · LESSON AREA

Computer vision for inspection and construction safety

09 · Computer vision for inspection and construction safety · AI for Civil Engineering

Course review
StandardInternational professional-learning synthesis · ABET 2026–27 · IEA GAPC v4 · ASCE ethics and AI responsibility · NIST AI RMF 1.0
Source3 source files
Review stateTechnical and publication gates pending
LEARNING OUTCOMES

After this chapter, you should be able to

  • Distinguish classification, detection and segmentation.
  • Audit image scale, labels and acquisition conditions.
  • Calculate detection metrics.
  • Assess domain shift and near-duplicate leakage.
  • Specify quality gates, abstention and inspector handoff.

Engineering context and methodSource §Lesson 09 · Engineering context and method · CASE-01 · RES-07 · RES-08 · DATA-07

An image records surface appearance under a particular scale, focus, lighting, angle and occlusion. A pixel, box or class does not establish depth, cause or structural consequence. Split by asset, site or acquisition campaign so near-duplicate crops cannot cross partitions. FHWA-HRT-24-055 describes research and validation activity; it is not evidence that AI replaces mandated inspection.

Illustrative inspection-triage visual. The schematic is not a crack-width measurement, condition rating or construction-safety decision.Original STRUCTURA review diagram · technical sign-off pending

Verified worked exampleSource §Lesson 09 · Verified worked example · CASE-01 · RES-07 · RES-08 · DATA-07

WORKED EXAMPLE

Image-screening confusion matrix

For 200 images, TP = 32, FP = 8, FN = 8 and TN = 152.

  1. Precision

    32 / (32 + 8)

    0.80
  2. Recall

    32 / (32 + 8)

    0.80
  3. Specificity

    152 / (152 + 8)

    0.95
  4. F1

    2×0.8×0.8/(0.8+0.8)

    0.80

Result. Eight labelled positives were missed. The system may rank images for review only, with capture-quality and out-of-domain abstention.

Practical lab · 5 h lesson effortSource §Lesson 09 · Practical lab · 5 h lesson effort · CASE-01 · RES-07 · RES-08 · DATA-07

  • Use the rights-cleared precomputed result table until image rights approval is recorded.
  • Compare thresholds and analyse at least five false positives and five false negatives.
  • Define blur, scale, lighting, occlusion and domain-shift quality gates.
  • Write an inspector handoff stating what the image output cannot determine.

Failure modes to investigateSource §Lesson 09 · Failure modes to investigate · CASE-01 · RES-07 · RES-08 · DATA-07

  • Near-duplicate crop leakage.
  • No physical scale or acquisition metadata.
  • Internet-image domain mismatch.
  • One annotator treated as unquestionable truth.
  • Segmentation described as calibrated crack width without evidence.

Knowledge checksSource §Lesson 09 · Knowledge checks · CASE-01 · RES-07 · RES-08 · DATA-07

Five review questions and answer rationales
QuestionAnswer rationale
Does a crack pixel imply structural distress?No; geometry, cause, depth and asset context require engineering assessment.
Why split by asset or campaign?Crops and conditions from one acquisition are correlated.
What triggers abstention?Inadequate quality, scale, view or out-of-domain content.
What is label uncertainty?Qualified reviewers may disagree or lack sufficient evidence.
What does the FHWA case establish?A research system is being developed and validated, not approved as replacement inspection.
Use these as formative checks. Technical and editorial review remain pending.

Key points

  • Start from the accountable engineering decision and its consequence.
  • Compare against a transparent non-AI baseline.
  • Validate on a split that represents intended use and retain human authority.

Source references recorded by the supplied chapter

  • FHWA-HRT-24-055, Artificial Intelligence for Bridge Inspection research report.
  • Peer-reviewed civil-infrastructure vision reviews listed as RES-07 and RES-08.
  • ASCE Policy Statement 573.