ninety percent false

alarms, workers stop listening

Problem statement

"Struck-by" incidents — workers hit by vehicles, cranes, or falling objects — are the second leading cause of construction fatalities in the U.S. (approximately 75–100 deaths annually) and cause thousands of serious injuries. Proximity warning systems (PWS) that alert equipment operators when workers are within a danger zone have been commercially available for over a decade, using technologies including radar, RFID, UWB radio, camera-AI, and magnetic field sensing. In controlled test environments, these systems achieve detection rates above 95%. On active construction sites, however, false alarm rates exceed 90% — workers and operators quickly learn to ignore alerts that fire constantly, rendering the systems worse than useless. The high false alarm rate is caused by the construction environment itself: constantly changing site geometry, multiple workers and machines in close proximity, reflective surfaces, metal structures, and dynamic work zones where "danger" and "normal operation" overlap spatially.

Why this matters

Beyond the 75–100 annual fatalities, OSHA estimates 10,000+ nonfatal struck-by injuries per year in U.S. construction, with direct costs exceeding $500 million. OSHA has considered mandating proximity detection on construction equipment (as mine safety regulations already require for underground mining) but has not issued a rule, partly because available technology cannot achieve acceptable false alarm rates in construction environments. The mining sector's success with proximity detection (mandated in multiple jurisdictions) demonstrates that the technology works in more controlled environments — the barrier is specific to construction's dynamic, unstructured worksite.

What’s been tried and why it hasn’t worked

Radar-based systems detect workers but cannot distinguish between a worker in the danger zone and a steel beam, concrete form, or reflective surface. RFID/UWB active tag systems require every worker to wear a tag — compliance rates in studies are 60–80%, and tags interfere with certain construction equipment. Camera-AI systems achieve good performance in clean environments but degrade severely in dust, rain, low light, and when workers are partially occluded by equipment or materials. Geofencing approaches require continuously updating digital boundaries as the site changes daily — an administrative burden that project managers don't maintain. The fundamental problem is that construction sites are not structured environments: the spatial relationship between workers, equipment, and hazard zones changes hourly, and current systems cannot distinguish "near the machine and working" from "near the machine and in danger."

What would unlock progress

Context-aware proximity detection that integrates worker position with equipment operational state (boom extending, vehicle reversing, load swinging) and task context (worker is in the designated rigging zone vs. wandering into the swing radius). This requires fusing real-time equipment telemetry with worker positioning and a dynamic model of the evolving worksite. The key insight is that danger is not proximity alone but the intersection of proximity + equipment motion + task state. Autonomous vehicle technology in construction equipment (already emerging) could provide the equipment-state data needed; the integration challenge is stitching together worker positioning, equipment telemetry, and site model in real time.

Entry points for student teams

A team could build a scaled surrogate of the problem on a campus lot — a rented skid-steer or an RC-scale excavator instrumented with position/orientation sensors — and test danger zones that change shape with the machine's operational state, comparing false alarm rates against a static-zone baseline without ever entering an active jobsite. A machine-learning team could work from the public ACID construction image dataset (acidb.net, 10,000 labeled images of construction machines) to quantify how worker-and-equipment detectors degrade under real-site clutter, dust, and occlusion — the perception failures that drive false alarms. Validating either result on an active site requires a contractor partner willing to carry the liability, OSHA credentialing, and worker-filming consent involved; that step is a follow-on handed to the partner, not the semester deliverable. Relevant disciplines: sensor fusion, robotics, construction management, human factors, machine learning.