Sleeping on Duty Detection: Guard & Operator Alertness

A guard post, a control room, a night-shift pumping station, a lone reception desk after midnight — each of these exists precisely so that someone is awake and available. When that person is asleep, the entire control that the role represents is absent, and nothing about the site looks different from the outside. Sleeping on duty detection addresses a gap that is otherwise almost impossible to manage: you cannot supervise someone whose whole job is to be present.

Attention and fatigue at work are well-studied occupational concerns; the U.S. Bureau of Labor Statistics tracks injury and fatigue-related data, and the architectural choice to run this analysis locally rather than in the cloud is part of the broader shift described in Edge computing.

Why the Role Exists, and Why It Fails

Alertness-dependent roles exist because the consequence of an unattended period is high: an unchallenged entry, an unobserved alarm, a process excursion nobody notices until morning. The failure mode is rarely misconduct. It is fatigue on a night shift, a long static watch with nothing to do, and a body that eventually rests.

• Security posts and gatehouses — where the whole value of the post is a person being awake.
• Control rooms — operator monitoring of alarms and process displays.
• Lone workers — night reception, remote pumping stations, and single-person industrial sites where a collapse or medical event is indistinguishable from sleep.
• Fire and safety watch — hot-work fire watch, confined-space attendants, and the roles where continuous observation is a formal control.

That last category is the important one: in several of these roles, staying awake is not a productivity expectation but a documented safety control, which means failing to verify it is a compliance exposure rather than a management preference.

What the System Actually Sees

Detection combines posture and duration rather than looking for a single tell-tale sign.

• Head and torso posture — a slumped posture with the head dropped forward or resting on a surface, sustained over time.
• Immobility — the absence of the small movements a waking person makes continuously. This is the strongest signal, and the hardest to fake.
• Eye closure where the view allows it — a close, frontal camera can read prolonged eye closure, though this needs a reasonably clear facial view and should not be relied on alone.
• Desk or workstation posture — head down on a desk or console for longer than a configured period.

The decisive parameter is duration. Everyone closes their eyes; nobody closes them for ninety seconds while upright. Requiring sustained immobility is what makes the detector reliable, and that threshold is set per site — longer for a quiet control room, shorter for a fire watch where the standard is formal.

Pairing With Related Detectors

On its own, a sleep alert tells you the post is unmanned. Combined with other detectors it starts to tell a fuller operational story. Pairing with phone use detection and looking at phone detection addresses the wider attention problem, of which sleep is only the most extreme form. Pairing with presence logic — confirming whether the post is occupied at all — distinguishes a sleeping guard from an abandoned one, which is a more serious event and needs a different response.

In manufacturing, the same cameras are usually worth configuring for factory safety rules, and in manufacturing operations the alertness question often extends to machine operators as well as security staff.

Processing runs on-device across 2–128 channels and 1–256 TOPS, with the full library of 198+ algorithms available on the same hardware. Local processing matters here for an obvious reason: an alertness monitoring system that uploads continuous video of staff to a third-party cloud is a very difficult system to defend to a workforce.

The longer explanation sits in elderly fall detection.

See it applied end to end in our guide to fatigue and driver state monitoring.

FAQ

Can it detect someone pretending to be awake? It detects sustained immobility and posture rather than intent, so briefly opening the eyes does not reset it. If a person remains motionless in a sleep posture beyond the configured duration, the alert fires regardless. What the system cannot do is judge alertness in an upright, moving person — that is not a claim any video system should make.

Is this the same as driver fatigue monitoring? The underlying idea is related but the implementation differs. In-cab driver monitoring works with a close, fixed, frontal view of one person and can read eye and head movement finely. A fixed post camera works at greater distance with more variable posture, so it relies more heavily on immobility and body position. They are related capabilities, not the same detector.

Do we need to install new cameras? Usually not — the analytics consume existing RTSP streams from an edge box. Check that the camera actually frames the workstation, since a camera pointed at a door will never see the desk. Because inference is local, monitoring continues with no internet connection and video stays on your premises.

What should we budget? A one-time hardware cost determined by channels and compute, with no per-camera subscription, so coverage can be extended later without a recurring charge. Given the sensitivity of the use case, most organisations run a Starter Kit trial and agree the policy with staff representatives before a full rollout.

Tell Us Your Scenario. Tell us which posts are alertness-critical and we will recommend camera framing and duration thresholds — and help you frame the policy before you switch it on. Start with a Starter Kit, or contact us to discuss your sites.