Vision-Based Safety Monitoring and Uncertainty-Aware Inference | Research

The problem

Real-time safety coverage on a large program is arithmetic: officers times the fronts each can watch, divided by the fronts running in parallel. On a hyperscale campus that ratio gets small, and the chance that a brief unsafe act is seen drops with it. Cameras and drones can hold coverage near one per front, but a model that cannot say how sure it is cannot be trusted with a crew, and a signal nobody acts on changes nothing.

The approach

Detection models trained on site imagery, UAV pose estimation, and weather-integrated risk scoring, each paired with calibrated uncertainty so the system sends low-confidence cases to a person instead of scoring them with false precision. The human review step is part of the design from the start, so the system supports the safety team rather than watching the crew.

Figure

What the Figure Shows

Real-time coverage falls as a program runs more workfronts in parallel; instrumented coverage holds. The gap is acute exposure no one is watching.

Research → Think Tank → Field Instruments

Papers, Notes, and Instruments

Papers evidence

Computer Vision for Safety Management in the Steel Industry2024
UAV-DL-Based Computer Vision Framework for Construction Worker Pose Estimation2025
Weather-Integrated Computer Vision Framework for Real-Time Construction Safety Risk Assessment2026
Information-Theoretic Uncertainty Quantification for Vision-Based Construction Ergonomic AssessmentIN PRESS
Reliability-Gated Ergonomic Inference (RGEI): Calibrated Uncertainty and Human-in-the-Loop Risk Screening in ConstructionUNDER REVIEW
Integrating Wearable Sensing Devices and Computer Vision for Safety Management in Steel Mills2024
Computer Vision for Pipeline Monitoring Using UAVs and Deep Learning2023

Think Tank notes explanation

Field Instruments use

Autonomous UAV Site Monitoring

Named frameworks

Active work

Projects in This Area

UNDER REVIEW

Reliability-Gated Ergonomic Inference (RGEI)

Journal manuscript

Calibrated uncertainty for vision-based ergonomic screening, so the system routes low-confidence cases to a human instead of scoring them with false precision.

What it needs

Site partners for field validation of the human-in-the-loop gate

FIELD VALIDATION

Weather-Integrated Computer Vision for Real-Time Safety Risk

Presented at CRC 2026, extending to site trials

Folds ambient conditions into vision-based hazard scoring so that the same posture or task reads differently at 95 degrees and 20 mph gusts than it does on a mild morning.

What it needs

Camera-equipped sites with on-site weather logging

IN DEVELOPMENT

Safety Copilots, Not Panopticons

Framework manuscript

A collaborative-intelligence framework for human-AI partnership in construction safety, built on the premise that monitoring workers is not the same as helping them.

What it needs

Co-authors in human factors and organizational behavior

Still open

Open Problems

Calibration that survives weather, lighting, and camera drift across a season on one site.
The routing rule: when a vision system should hand a case to a human, and how to measure whether it did so at the right moment.
Deployments workers accept: what a safety copilot must show, withhold, and explain to be trusted on the workfront.

Research in motion

Recent in This Area

June 2026
Information-theoretic uncertainty quantification accepted at ISARC 2026Calibrated uncertainty for vision-based ergonomic assessment, the technical basis for routing low-confidence cases to a person.

Get in touch

Work With the Lab on This Area

Co-investigators in computer vision, human factors, and occupational safety; contractors or industrial operators with camera or UAV access on a live site.

Propose a Collaboration Offer a Site or Dataset Student Research

Previous area · 04Project Delivery Outcomes and Construction Education Next area · 02Ergonomic and Physiological Exposure at Height and in Heat