Driver Fatigue Detection: Video Signs and Site Deployment

Driver fatigue detection is the video analytic that watches for the moment before an incident rather than after it. Where most site analytics react to something that has already happened – a person inside a fence, a vehicle in a fire lane – fatigue detection looks for a physical state: eyes closing for too long, a head dropping forward, a jaw opening in a yawn. The alert arrives while the driver is still able to respond, which is the entire point.

This guide covers what the algorithm actually measures, the two very different ways it gets deployed, where accuracy breaks down, and what it takes to run on infrastructure you already own.

What Fatigue Looks Like on Video

Fatigue is not directly observable – the algorithm infers it from a cluster of signals that collectively distinguish a tired person from a bored one. The common indicators:

• Eyelid closure duration – blink rate rises with fatigue, and the fraction of time the eyes are shut grows. A blink that lasts noticeably longer than normal is the single strongest signal.
• Head pose – a slow forward nod, or a head that droops and is corrected, follows a distinct pattern that differs from looking down at a screen.
• Yawning – repeated wide mouth opening, usually combined with the two above rather than used alone.
• Micro-corrections – small steering or posture adjustments made to stay in a position, which show up as restlessness.
• Gaze stability – attention drifting off the forward view for longer than the surrounding behaviour would explain.

The design principle is that no single signal is reliable, so the algorithm combines them over a time window. This is why the output is usually a graded state rather than a binary alarm: a rising fatigue score raises a warning, and a sustained high score escalates it.

Why It Matters Operationally

Fatigue sits at the intersection of safety and liability, which is why it attracts attention from more than one department. A tired driver damages vehicles, cargo and infrastructure, and the resulting investigation asks a question that footage alone answers poorly: what was the driver’s state in the minutes before. A detection layer turns that into a timestamped event.

It is also a workplace health question rather than only a security one. Fatigue is managed as an occupational risk in most jurisdictions, and treating it that way changes who owns the system – not only the security team but the operations and safety functions. For reference on fatigue as a workplace hazard, see the UK Health and Safety Executive.

Two Deployment Shapes – and Why the Difference Matters

This is the part most write-ups skip, and it is the part that decides whether a project is straightforward or not.

In-vehicle detection places a camera facing the driver inside the cab. It sees faces at close range and consistent angles, which makes the analysis easier – but it requires hardware in every vehicle, a driver-facing camera position, and connectivity to move events off the vehicle. It is a fleet-technology project.

Fixed-point detection uses cameras that are already on the site: at a gatehouse, a loading bay, a weighbridge, a yard exit, or along an internal road. Here the driver is seen through a windscreen from outside, at a distance, under whatever light the site provides. That is a harder image problem, but it requires no vehicle hardware at all.

The honest position is that these are different problems with different answers. A site that wants in-vehicle monitoring is asking for a fleet deployment. A site that wants to know when a contractor’s driver has been stationary at a bay for an unusually long time at 03:00 is asking a fixed-camera question, and that is the one an edge AI box on existing streams can address.

Fatigue Is Not Sleeping on Duty

Related detections get bundled together, but they answer different questions. Sleeping on duty detection covers a stationary post – a guard or operator who has stopped moving and closed their eyes. Fatigue detection covers a person who is still nominally active but whose physical state is deteriorating, which is a much earlier and more useful warning.

It also overlaps with distraction rather than fatigue. Looking-at-phone detection and the broader phone use detection rules catch attention diverted to a device, while workstation abandonment detection catches the opposite failure – a position left unattended. Sites with vehicle operations usually find fatigue and distraction rules are reviewed in the same queue, because both produce “driver was not fit to be moving that vehicle” events.

Where Accuracy Breaks Down

Video-based fatigue inference degrades predictably, and pretending otherwise wastes everyone’s time:

• Distance and windscreen glass. From outside a cab, the face occupies few pixels and reflections sit between the camera and the eyes. Close-up in-cab views are simply easier to analyse than long-range exterior views.
• Eyewear. Sunglasses and strong prescription lenses obscure the eyelid signal, which is the strongest indicator. Sites should expect reduced confidence here rather than assume it works.
• Lighting. A cab in shadow, a driver backlit by yard floodlights, or headlights washing out the windscreen all degrade the image. Fixed lighting for the camera position is worth more than parameter tuning.
• Genuine ambiguity. A long blink and a glance down at a dashboard look similar in a single frame. Time-window logic separates them most of the time, not always.
• Cultural and individual variation. Expressive behaviour differs between people, so a threshold tuned on one workforce can misfire on another.

The practical consequence: expect a commissioning period, and expect the rule to be tuned to your site rather than delivered working out of the box. That is the same pattern as any behavioural analytic – see how to reduce false alarms for the general method.

Deployment on Existing Infrastructure

For the fixed-point case, nothing about fatigue detection requires new cameras on the site perimeter or in the yard. An edge AI box ingests the existing RTSP or ONVIF streams and runs the models on-device. Systems scale from 2 to 128 channels with 1 to 256 TOPS matched to the channel count, and the same appliance carries a library of 198+ pre-built algorithms – so a site that starts with fatigue rules at the gatehouse can add vehicle detection, speed or PPE rules on the same hardware later.

On-device processing also settles the data question. Driver footage is sensitive for both privacy and employment reasons; keeping analysis on site means video is not streamed to a vendor cloud to be processed there. Where the streams run over an existing camera network, the streaming layer is standard – see IETF RFC 2326 (RTSP) for the protocol that carries them.

Fleet-adjacent sites tend to combine this with wider yard rules: logistics operations pair fatigue and speed detections, while transportation sites usually add them to an existing depot-wide analytics layer. Sites operating hazardous materials add industrial and energy safety rules alongside, since driver state and load handling are reviewed by the same supervisor.

Where a site needs a detection type that is not in the standard library – an unusual camera geometry, a specialised vehicle type, or an in-cab view – it becomes a custom model scoped against your footage. That is why the first step is sharing video rather than agreeing a deadline, and why bespoke development carries no committed date until the footage has been reviewed.

Common Questions About Driver Fatigue Detection

Does this require cameras inside every vehicle? Not for the fixed-point case. Yard, gate and loading-bay cameras that already exist can carry the rule. In-vehicle monitoring is a separate fleet deployment requiring hardware in each cab, and it is a different project with different costs.

Can it work through a windscreen? It can, with reduced confidence. Distance, glass reflection and lighting all sit between the camera and the driver’s face. A camera position that looks through the side window at a consistent angle, with controlled lighting, produces materially better results than one firing down a long approach road.

What happens when the driver wears sunglasses? Confidence drops, because eyelid closure is the strongest signal and it is hidden. The rule can still use head pose and longer-window behaviour, but sites should regard sunglasses as a known limitation rather than expect equivalent performance.

Is the footage stored in the cloud? Not with an on-premise deployment. The appliance sits on site, analysis happens locally, and footage does not need to leave the premises to be processed. Retention and access policy remain a site decision, and worth writing down before go-live.

Next Step

Tell us where the drivers you care about are seen – gatehouse, loading bay, yard exit – and share a short clip from those cameras. We will tell you what a fatigue rule can realistically do at that angle, and what the configuration would look like. Order a Starter Kit to validate on your own footage, or tell us your scenario for a configuration quoted per site.