Everything works until the cameras switch to infrared at dusk

The collapse is caused by the camera, not by the darkness. At a lux threshold set in the camera's own configuration, a mechanical infrared cut filter swings out of the light path and the sensor starts seeing near-infrared. The frame becomes monochrome, the illuminator floods the scene from beside the lens, the exposure lengthens and the gain climbs. A model trained on daytime colour frames is now being asked to score an image type it has never seen, and it does exactly what you would expect.

That is why the first move is a chart, and why the usual version of the chart is misleading. Detections plotted against sunset show a smeared decline across 90 minutes, which invites theories about ambient light. Detections plotted against each camera's own switch show a step. Steps have single causes; smears do not. Getting this right early saves a fortnight of model work that would not have helped, and it applies across most of safety, PPE and site monitoring.

Finding the exact moment each camera switched

Switch times differ per camera and drift over months. A camera under a sodium lamp switches late or oscillates; one behind a dusty dome switches early; one facing a bay door switches twice as a vehicle's headlights sweep past. Asking the VMS for a nominal time is not good enough at that resolution.

1. Detect the switch from the frames themselves. Convert 1 sampled frame per minute to HSV and take the mean saturation across the image. In colour mode a working scene sits well clear of zero; the moment the filter drops out the mean falls to within 1 or 2 points of zero and stays there. That crossing is the switch, timestamped from the frame rather than from a log.
2. Bucket detections into 5-minute bins for the 4 hours around it, per camera and per rule, and align every camera to its own crossing rather than to the clock.
3. Separate the counts by rule class. A person-presence rule and a vest-compliance rule usually fall by very different amounts, and the ratio between the 2 is the diagnosis.
4. Pull 20 clips from either side of the crossing on the worst camera and look at them. Everything after this point is confirmation of something you can see in a clip, and teams that skip this step spend weeks on the wrong cause.

Why a hi-vis vest stops being hi-vis

2 separate things happen to the same garment, and they push in opposite directions. First, the colour goes. An orange vest and a grey jacket become two similar mid-tones, so any class whose separability rested on hue has lost most of its signal. A detector that appeared to be finding vests may in fact have been finding orange, and dusk is when that becomes measurable.

Second, the retroreflective tape on the vest does what it was engineered to do: it returns light straight back towards its source. With the illuminator mounted around the lens, the tape returns almost all of it into the sensor and saturates. What the model receives is a bright white bar across a dark torso, with the garment either side of it crushed into black. That is not a degraded version of the daytime appearance; it is a different object. Helmet stripes and reflective trouser bands behave the same way, which is why a helmet rule can fail at night while a person rule does not.

In infrared, a high-visibility vest is not a bright orange garment. It is a white bar on a black shape, and no daytime example looks like that.

The four other things the switch changes

The last row is the one that hides best. Some cameras are configured, sometimes by default, to change the sub-stream at night while the recorded main stream stays as it was. Reviewers pull the recording, see acceptable footage, and conclude the analytics are at fault — when the analytics were handed something else entirely. Check the profile the analytics pipeline is actually consuming, at night, on the camera, not in the design document.

Fixes, ranked by what they cost and what they actually buy

• Add white lighting to the zone. The most effective fix by a distance, because it keeps the camera in colour mode and the entire daytime model applies unchanged. It is a facilities job rather than a software one, it changes how the area feels to work in, and in a yard or a loading bay it is often already justified on safety grounds alone.
• Move the illuminator off-axis. Mounting the infrared source 1 metre from the lens rather than around it removes most of the retroreflective bloom, because the tape no longer returns light along the viewing axis. Cheap, physical, and frequently enough on its own.
• Fit an infrared-corrected lens. If the whole frame is soft at night and sharp by day, the lens is focusing visible and infrared light at different planes. No training set fixes this, and no other fix works until it is done.
• Cap the exposure time and accept the noise. A maximum shutter of 1/100 second trades grain for sharpness, and for moving-person rules that is usually the right trade — but it makes the far half of the scene worse, so verify per camera rather than fleet-wide.
• Build a separate night training set. Label infrared frames as their own data and either train a second model or fine-tune with them well represented; do not simply pour them into the daytime set, where they are outnumbered and ignored. Whether the night problem is better posed as named classes or as a model of what normal looks like is the same choice argued in anomaly detection versus supervised defect classification.
• Re-specify the rule for darkness. Detect person plus helmet shape rather than vest colour, or fall back to presence and zone rules that need only a body-shaped region. Less capable, honestly labelled, and available immediately.

One lever that looks tempting and is not: lowering the detector's confidence threshold after dark. It recovers the count on a dashboard and does it by converting silent misses into noisy false alarms, which is the fastest known way to produce an alert queue nobody works — the collapse described in three hundred alerts a shift and nobody opens them.

The rules that should simply be switched off after dark

Some rules cannot be run honestly on infrared frames with the cameras a site actually has, and continuing to run them is worse than stopping. A colour-dependent rule that quietly misses 8 violations in 10 between 19:00 and 06:00 does not produce a gap in the record; it produces a clean record, and a clean record is read by everyone as compliance. That is an active harm, not a limitation.

So state it explicitly: a per-rule schedule, visible on the operator screen, showing which rules are live in which lighting mode. The night shift is entitled to know that vest checking is not running and intrusion detection is. Coverage should be re-audited under night conditions too, because an illuminator's usable range is far shorter than the lens's daylight field of view, and a zone that is covered at noon can be effectively uncovered at midnight — the audit method is in the incident nobody has footage of.

Proving the fix before the next night shift trusts it

Hold out a labelled night set before changing anything: 300 to 500 frames per zone drawn from after each camera's switch, covering the movements you care about. Score every candidate fix on that set alone. A mixed day-and-night evaluation is dominated by daylight frames and will show a comfortable overall figure while night recall stays where it was, which is how a fix gets signed off and the problem persists.

If the answer turns out to be the cameras rather than the model — soft lenses, on-axis illuminators, sub-streams that change profile — that is a procurement finding, and it belongs in the estate review rather than the backlog. Sorting an inherited camera estate into reusable, repositionable and replace is the exercise in can the cameras you already own do this. Doing that survey once, with night performance included, is cheaper than the second round of model work — and it is the kind of unglamorous groundwork that decides whether a manufacturing and industrial vision deployment survives its first winter. Where the answer is genuinely a build, scope it as a product build with the night set as an acceptance criterion rather than as a later phase.

• video analytics
• PPE detection
• cameras
• night shift