Troubleshooting
Why your video analytics false-alerts all night and goes quiet by morning
If the alerts pile up between dusk and dawn and stop at sunrise, the model is usually fine. Something physical in front of the lens changed.
VuePoint
A camera that behaves all day and generates forty alerts a night is not a model problem, and retraining it will not help. Something in the physical scene changes at dusk — the illuminator comes on, the exposure opens up, the aperture widens, and the camera starts seeing things it does not see in daylight. Work the physical causes first. They account for most of it, and they are cheaper to fix.
Confirm it's actually a night pattern
Before changing anything, plot the alerts by hour for the last seven days. You are looking for a distribution, not an anecdote. Three shapes tell you three different things:
| Pattern | What it usually means |
|---|---|
| Sharp rise within ~20 minutes of sunset, sharp fall at sunrise | The IR illuminator. Insects, or the light itself hitting something reflective. |
| Spread across the night with clusters | Traffic-driven — headlights sweeping the scene, or a road/parking area in frame. |
| Starts a few hours after dark, worse on some nights than others | Weather. Rain, fog, and snow all scatter IR straight back into the lens. |
The physical checks, in the order worth running them
- 1
Look at the actual clips, at full resolution
Not the thumbnails. Open five or six of the false alerts and watch what triggered them. In most cases you will see the cause directly — a moth crossing 10cm from the lens looks like a person crossing the yard, because it is out of focus and moving fast.
- 2
Check for insects and spider webs at the housing
IR illuminators attract insects, insects attract spiders, and a web strung across the housing catches every one of them. A web moving in the wind, lit from behind by the illuminator, is the single most common cause of overnight false alerts. Clean the housing and the glass, and check again in a week — the web comes back.
- 3
Separate the illuminator from the camera if you can
The insect problem is caused by the light being co-located with the lens. An external illuminator mounted two or three metres away still lights the scene, but the insects gather around it instead of directly in front of the sensor. This one change fixes more overnight alerting than any software setting.
- 4
Look for retroreflective surfaces in frame
Road signs, high-vis vests on a rack, licence plates, reflective tape on a bollard, and wet painted lines all bounce IR back hard. The camera reads a blown-out bright patch that appears and disappears as the exposure hunts. Re-aim slightly, or mask the region.
- 5
Check whether the exposure is hunting
Watch a clip and look at the overall image brightness. If it pulses — bright, dark, bright — the auto-exposure is oscillating, usually because a bright light source is at the edge of frame. Lock the exposure for the night profile, or move the light source out of frame.
- 6
Check the day/night switch itself
A camera sitting right on the threshold will flip between colour and IR modes repeatedly at dusk and dawn. Every flip is a whole-frame change. Widen the switching hysteresis, or set a fixed schedule instead of a light-level trigger.
- 7
Only now, look at detection settings
If the scene is clean and the alerts continue, then tune: minimum object size, dwell time before alerting, and detection zones that exclude the sky, the road, and anything beyond the property line. Raise the minimum size until an adult at the far edge of the area you care about still triggers, and nothing smaller does.
Weather, and what you can't fix
Rain and snow close to the lens are lit by the illuminator and scatter light back — the image fills with bright moving specks and every motion-based trigger fires. Fog does the same thing more evenly. There is no setting that removes this while leaving real detections intact, because to the sensor they are the same photons.
- A hood or extended weather shield keeps precipitation out of the near field where the illuminator lights it brightest.
- Angling the camera slightly downward reduces the volume of lit airspace in frame.
- Separating the illuminator from the lens helps here too, for the same reason it helps with insects.
- Detection that requires an object to persist and move coherently across several seconds is far more robust to precipitation than frame-differencing motion.
If none of that changes anything
Then it is worth looking at the detection stack. Capture a representative night's clips, note what fired and what it should have said instead, and take that to whoever runs your analytics. A specific list of twenty misfires with timestamps is worth more than a description of the problem — it is the difference between a support ticket that gets resolved and one that gets a settings suggestion. If you want a second opinion on a set of clips, send us the details — it does not need to be our hardware.
Common questions
- Will switching to a different analytics model fix night-time false alerts?
- Usually not on its own. If a spider web is moving 10cm from the lens, every model sees a large fast-moving object, because that is what is in the frame. Fix the scene first; then the model choice matters.
- Is it better to use white light instead of IR?
- White light avoids the IR-specific problems and gives you colour at night, which makes evidence far more useful. It also attracts insects, is visible to anyone on site, and is not always permitted — check local lighting rules and neighbouring properties before switching.
- How many alerts a night is normal?
- There is no universal number — it depends entirely on what is in frame and what you have asked the system to tell you about. The useful question is what proportion of alerts you act on. If it is under about one in ten, the configuration needs work, not your attention span.
Still getting woken up?
Send us a night's worth of clips and what you expected the system to say. We'll tell you what we'd check next — it doesn't have to be our equipment.