The platform, in detail
HardHat Pulse runs a full detection stack — from edge inference on RTSP streams to OSHA 300 log generation — without requiring internet connectivity for real-time alerts.
How HardHat Pulse works
The detection pipeline runs on an on-site edge server. No video leaves the jobsite. Alerts reach supervisors in under 1.2 seconds.
Ingest & Stream
Up to 48 RTSP camera streams feed into the edge inference node. The node decodes each feed at 10fps for detection and archives 30fps video for incident review. Wearable sensor beacons transmit over a local 900MHz mesh — independent of site WiFi.
Detect & Classify
A fine-tuned YOLOv8 model runs per-frame inference to detect hard hats, high-visibility vests, safety glasses, and gloves. Worker silhouettes are tracked across cameras using re-identification embeddings, so an alert follows the person — not just the pixel location.
Alert & Log
Detections crossing configured confidence thresholds generate alerts via push notification, SMS gateway, or site PA integration. Every event writes to a tamper-evident SQLite archive with frame hash, GPS coordinate, worker tag ID, and detection confidence score.
Real-time site overview in one screen
The supervisor dashboard shows live camera thumbnails, active alert count, zone occupancy, and worker count — updated every second. Drill into any camera stream or click an alert to see the frame-level evidence that triggered it.
Alert history exports as PDF or CSV. OSHA 300 log entries auto-populate from the detection database — no manual data entry required at the end of the month.
What the system actually does
Each capability addresses a specific failure mode in traditional construction safety programs.
Multi-Class PPE Detection
The detection model classifies 14 distinct PPE categories including ANSI Class 2 and Class 3 vests, Type I/II/E hard hats, ANSI Z87.1 safety glasses, and cut-resistant gloves. Detection confidence thresholds are configurable per PPE class — high-consequence items like hard hats default to 0.82 confidence, lower-risk items to 0.70.
False positive rate stays below 3.1% in tested deployments. The model was trained and validated on labeled footage from active U.S. and Canadian jobsites — not studio simulations.
Geo-Fenced Zone Management
Safety managers draw polygon exclusion zones directly on the site map within the web interface. Zone definitions tie to specific time windows — an exclusion zone can be active during concrete pours and automatically suspended when the foreman clears it.
Worker entry into a restricted zone triggers a three-stage alert: first to the worker's wearable haptic band (immediate vibration), then to the site foreman app (push notification), then to the site PA system (audio announcement) if the worker doesn't clear the zone within 15 seconds.
Wearable Physiological Monitoring
The HardHat Pulse sensor clip attaches to any standard hard hat brim. It logs heart rate, skin temperature, and 3-axis acceleration at 50Hz. On-device inference detects falls (sudden acceleration spike followed by stillness), heat stress (sustained elevated heart rate + temperature), and slip events.
Battery life is 14 hours on a single charge. Clips charge via USB-C in the site trailer. The edge gateway maintains a local mesh network — alerts fire even without cellular coverage inside underground or enclosed jobsite areas.
Camera Management & Playback
The platform manages fixed RTSP cameras (supported: Axis, Hikvision, Dahua, Hanwha), PTZ cameras with auto-pan to alert coordinates, and optional helmet-mounted GoPro-class cameras via the companion mobile app. Camera health monitoring reports dropped frames, lens obstructions, and connectivity outages to the site supervisor.
Incident review mode shows 30-second pre-event and post-event clips with detection overlays burned in — the same frame evidence that feeds into the OSHA log entry.
Safety Trend Analytics
Weekly and monthly reports show alert frequency by zone, PPE violation rates by shift, and near-miss incident patterns. The analytics engine correlates alert density with schedule data from Procore — identifying high-risk periods like concrete pours, steel erection, and framing days before they start.
GC and owner stakeholders get a read-only dashboard URL. No separate account required. Safety consultants can pull raw CSV exports for third-party analysis.
API & Webhook Integration
The REST API supports alert subscriptions, detection log queries, zone management, and worker tag registration. Webhooks deliver real-time alert payloads to any HTTPS endpoint — including Procore Daily Logs, Slack channels, and custom EHS management systems.
OpenAPI 3.0 spec is publicly available. Procore and Autodesk ACC integrations install from their respective app marketplaces. Average integration setup time for a new jobsite is 4 hours including camera provisioning.
Performance benchmarks
Measured on production deployments. Single-site averages across 12 active pilot installations.
Detection Latency
<1.2sCamera frame to supervisor alert, measured end-to-end including edge inference and push delivery
PPE Accuracy (F1)
97.3%F1 score across all 14 PPE classes on the held-out site validation set
Camera Capacity
48Simultaneous RTSP streams per edge node on a single NVIDIA Jetson AGX Orin
Sensor Accuracy
99.1%Fall detection true positive rate, measured in staged-fall hardware testing
Video Archive
90 days30fps full-resolution event archive retention on local NVR; alert clips replicated to cloud
Sensor Battery
14 hrsHard hat sensor clip runtime per charge; bulk USB-C charging dock supports 50 clips
Ready to instrument your site?
Schedule a 30-minute technical demo. We'll walk through the full detection stack and answer your integration questions.
Schedule a Demo