AI Generates Synthetic Highway Construction Hazard Images from OSHA Reports
A recent study published on arXiv (2605.11276) introduces a generative AI technique designed to produce synthetic visual representations of highway construction hazards based on narratives from OSHA Severe Injury Reports. This method generates images in a single pass (one for each incident) and creates four-stage temporal sequences. Utilizing 75 incident records, the researchers generated 750 images, which were assessed through CLIP-based semantic retrieval and expert evaluations. The single-pass images received an educational acceptability rating of 81.1%, with fidelity and alignment scores of 4.14/5 and 4.07/5, respectively. Meanwhile, the temporal sequences garnered a 60.9% acceptability rate. This research aims to tackle the lack of image-based training resources stemming from ethical and logistical challenges.
Key facts
- arXiv paper 2605.11276
- Generative AI methodology for highway construction hazard visualization
- Uses OSHA Severe Injury Report narratives
- Two modes: single-pass and temporal four-stage sequences
- 75 incident records produced 750 images
- Evaluated with CLIP-based semantic retrieval and expert assessment
- Single-pass images: 81.1% educational acceptability, fidelity 4.14/5, alignment 4.07/5
- Temporal sequences: 60.9% acceptability
Entities
Institutions
- arXiv
- OSHA