A narrow set of construction safety AI use cases holds up on an Australian site. This page lists those, and the ones that do not.
Effective approaches
- PPE detection at the gate. Hard hat, hi-vis, glasses. Real-time, edge-based. Flags before entry, not after.
- Site density alerts. Aggregate counts per zone. Triggers an operations alert when a zone exceeds planned occupancy.
- Vehicle classification at the gate. Forklift vs truck vs car. Anonymous, no ANPR.
Current limitations
- Worker-level tracking. Identifying "John Smith" and logging his location across the site. Legally and culturally uncomfortable in Australia.
- Emotion or fatigue detection. The models are not reliable enough, and the regulatory risk is high.
- Helmet cameras streaming to the cloud. Bandwidth, cost, and the WHS team's willingness to sign off.
Getting started
Start with one gate, one PPE model, one week of data. The WHS team then has near-miss counts by hour and by gate, and the system has a calibration baseline. If the counts are useful, expand to the next gate.