How Vision-Based PPE Detection Is Improving Worker Safety on Construction Sites

The Challenge

Safety on each site depended on walk-arounds by a small number of safety managers who also handled toolbox talks, paperwork, and incident reports.

The existing approach made it difficult to:

• Observe PPE compliance across large, changing sites throughout a full shift.
• Apply zone-specific rules, such as fall protection near open edges or eye protection near cutting work.
• Respond to safety gaps while crews were still in the zone.
• Spot patterns by time of day, trade, or area before they led to incidents.
• Give subcontractors objective feedback rather than anecdote.
• Produce consistent records for safety reviews and audits.

From a technical perspective, construction footage is hard for vision models. Workers appear small and partly hidden behind scaffolding and materials, lighting shifts from dawn to floodlights, and dust and rain degrade images. Site connectivity was patchy, so most processing had to run on local hardware, and the system had to monitor safety behavior without tracking individuals.

The Solution

InnovationM developed an edge-first PPE detection system that reads existing site camera feeds, applies rules per zone, and sends supervisors short, actionable alerts. Instead of recording and identifying people, it focuses on whether the right equipment is present in the right place. The solution included:

• Camera onboarding: Integrating the contractor's existing IP cameras over RTSP, with a review of placement, angles, and coverage gaps for each site.
• Person & PPE detection: Fine-tuning a YOLO-family deep learning object detection model on construction imagery to detect people, hard hats, high-visibility vests, safety glasses, and harnesses.
• Zone-based rules: Letting safety managers draw zones on camera views and set the PPE required in each, updated as the site changes.
• Edge inference: Running optimized models on compact GPU devices at each site, so footage stays local and alerts work even when the connection drops.
• Alert tuning: Using persistence checks across several frames and agreed confidence thresholds, validated through testing services, so a briefly hidden helmet does not trigger an alarm.
• Privacy by design: Blurring faces in stored snapshots, keeping short retention windows, and applying role-based access, with no facial recognition or individual identification.
• Safety dashboard: Showing trends by site, zone, trade, and time of day, with snapshots attached to each event for toolbox talks and audits.

The key focus was not simply detecting missing hard hats, but creating a timely, trusted signal supervisors could act on while the crew was still in the zone.

The Impact

The implementation helped the contractor move from occasional walk-arounds toward continuous, zone-aware PPE monitoring. Supervisors now receive a focused list of events rather than watching camera walls. The solution supported:

• Better PPE compliance in high-risk zones.
• Faster supervisor response to safety gaps.
• Less safety manager time spent on routine checks.
• Objective, site-level data to share with subcontractors.
• Consistent, timestamped records for safety reviews and audits.
• A foundation for broader computer vision safety monitoring across sites.

By bringing together existing site cameras, fine-tuned object detection, zone-aware rules, privacy-conscious edge deployment, and dependable data engineering, InnovationM showed how computer vision development services can give safety teams a wider view of every site without adding headcount.