Construction safety AI that knows what it doesn't see: a

Construction safety AI is getting closer to reliable PPE monitoring, but a persistent challenge remains: computer vision systems are often good at recognizing what’s there and weak at confirming what’s missing. In practice, that leads to two common failure modes—confusing background clutter with helmets or vests, and flagging workers whose PPE is only occluded, blurred, or too small to resolve. A new Frontiers in Built Environment study by Bhargav Vaghani of Arizona State University tackles absence detection with a two-tier “Sentry-Judge” design. The Sentry uses a YOLOv11m detector to quickly screen frames and assign uncertainty scores to provisional PPE-missing flags. Only when uncertainty crosses a threshold does the system trigger a SAM 3 segmentation model for more computationally intensive verification of worker bodies and PPE masks.