Fall detection from a head-worn sensor is a solved problem in elderly care — smartwatches and personal emergency response systems have been doing it accurately for a decade. The algorithms are well-understood. The problem in construction is not the algorithm; it's that construction workers do dozens of things per hour that look exactly like a fall to a 3-axis accelerometer, and the false positive rate in that environment is the thing that will determine whether your supervisors respond to fall alerts or start ignoring them.
A 3-axis accelerometer measures acceleration in X, Y, and Z axes at sampling rates of 100-400 Hz on current wearable hardware. A fall from height produces a recognizable signature: free-fall phase (near-zero gravity reading on all axes), rapid deceleration spike at impact (4-16G depending on fall height and landing surface), and post-impact stillness or erratic movement pattern. Distinguishing that signature from other high-acceleration events — which happen constantly on a construction jobsite — is where the engineering challenge lives.
Construction activities that produce fall-like signatures
Through 14 weeks of validated testing at our Houston pilot, we documented the construction activities most frequently associated with fall detection false positives. In rough order of frequency:
Heavy load set-down: placing a 50+ pound object on a surface — concrete block, steel plate, tool bag — produces a rapid deceleration signature when the worker bends forward and releases the load. The hard hat clip, positioned at head height, registers the forward rotation and deceleration. This is the most frequent false positive source, accounting for approximately 40% of all false positive events in our pilot data.
Aggressive bending and crouching: workers in constrained spaces frequently move between standing and crouching postures quickly. The transition from standing to crouching is kinematically similar to the initial phase of a forward fall — forward torso rotation with deceleration at the bottom of the movement arc.
Tool strikes and vibration: jackhammer operation, core drilling, and compaction plate operation all produce sustained high-amplitude vibration that can trigger threshold-based fall detection if the detection algorithm isn't designed to distinguish sustained periodic vibration from impact events.
Dropping the hard hat: workers who set their hard hat on a surface or drop it during a task generate impact events on the clip sensor that look, in isolation, like an impact event at the end of a fall. Context — specifically, that the badge is no longer moving after the impact rather than showing irregular post-fall movement — helps distinguish this, but not in all cases.
The detection algorithm design
Our fall detection algorithm uses a three-phase validation approach. Phase 1: detection of a free-fall event (acceleration below 0.3G on the combined vector for 200ms or longer). Phase 2: detection of an impact event (combined acceleration above 4G within 3 seconds of phase 1). Phase 3: post-impact analysis — is the subsequent motion pattern consistent with a fall (irregular struggling or stillness) or a normal work movement (immediate return to upright posture and continued work motion)?
Phase 3 is where most false positives are eliminated. When a worker bends to set down a load and the sensor fires phase 1 and phase 2 triggers, the phase 3 analysis sees the worker immediately returning to an upright walking posture and cancels the fall classification. When a genuine fall occurs, phase 3 sees either sustained stillness (worker incapacitated) or erratic movement without an organized return to upright posture.
The phase 3 window is 8 seconds from impact. A fall alert fires only when all three phases are confirmed. This three-phase approach reduces false positives by approximately 85% compared to a threshold-only detection approach, at the cost of 200ms additional alert latency (the phase 3 window adds time).
Measured false positive rates by work type
Across the 14-week Houston pilot with 87 wearable badges deployed, we measured 3.2 confirmed false positive fall alerts per worker per shift — the figure we publish in our product documentation. That breaks down by work type: concrete workers (most heavy load set-down activity): 4.8 per shift. Finish carpenters and interior workers: 1.9 per shift. Equipment operators (seated in cab, minimal sensor activity): 0.4 per shift.
Supervisors managing 100 workers on a mixed crew receive approximately 320 false fall alerts per shift. That sounds alarming until you understand the alert hierarchy: genuine fall alerts carry a distinct mobile notification type (red banner, continuous vibration) that supervisors learn to treat as immediate-response. The more frequent false fall candidates arrive as lower-priority events that are automatically dismissed if the worker's BLE position shows continued normal movement within 15 seconds of the candidate event.
The auto-dismiss feature eliminates approximately 78% of false positive fall candidates before they reach the supervisor's notification queue. The remaining 22% — about 70 per shift across a 100-worker site — require supervisor acknowledgment. That's still significant and is the subject of ongoing algorithm improvement.
Fall from height vs. same-level fall
Falls from height and same-level falls have different detection profiles. A fall from scaffold height (10+ feet) produces a free-fall phase of 700-800ms before impact — well above our 200ms threshold, generating a clear phase 1 signal. A same-level slip or stumble may produce only 50-100ms of reduced gravity before impact, which our current threshold treats as insufficient for phase 1 confirmation.
This means our current algorithm is more reliable for fall-from-height detection than for same-level trip-and-fall detection. Same-level falls are less likely to be fatal but more frequent and more likely to cause lost-time injuries (ankle fractures, knee injuries, wrist injuries from breaking a fall). The phase 1 threshold is configurable; some customers lower it to 150ms to increase same-level fall sensitivity, accepting a higher false positive rate in exchange for better coverage of slip-and-fall events.
We recommend the 200ms default threshold for most deployments and suggest lower thresholds only for sites where same-level slip-and-fall risk is elevated — specifically sites with wet concrete floors, active chemical spills, or elevated slip hazard ratings from the safety assessment.
Integration with the broader fall protection compliance system
Fall detection from the wearable sensor operates in parallel with camera-based fall protection compliance detection — workers observed near unguarded edges without visible lanyards or harnesses. The two systems address different scenarios. Camera detection handles the pre-fall compliance check: is the worker protected before they reach the edge? Sensor detection handles the post-fall event: did a fall happen, and does the site need emergency response?
The combination is more effective than either alone. A site that has both systems active can close the notification loop: camera alerts when a worker approaches an unguarded edge without visible fall protection (preventive), and sensor detection confirms if the prevention was insufficient and a fall occurred (emergency response). As discussed in the PPE detection accuracy article, camera-based compliance detection under adverse conditions is less reliable — the sensor layer provides continuity when camera performance is degraded.
Questions about fall detection tuning for specific work types or environments: contact us at contact@hardhatpulse.com. The algorithm parameters we've described here are the defaults; adjustments for specific site conditions are supported without a software release cycle.