Arlo camera false triggers, and every other cause of false alerts

Your camera is not broken and it is not lying to you. It is doing exactly what it was built to do, which is to report that a lot of pixels changed. The problem is that at two in the morning in November, a great many things change pixels and almost none of them are a person.

Here is each cause, why it happens physically, and what actually removes it. The list is in rough order of how often it is the culprit on a real site. If you are here about Arlo camera false triggers, or any other battery camera, start there instead: those wake on moving heat rather than on changing pixels, so they trip on different things and most of the fixes below only half apply.

What your camera is actually doing

Classic motion detection compares each frame against the last one, or against a rolling average of recent frames. It subtracts one from the other, counts pixels whose brightness changed by more than a threshold, and if enough of them are close together inside your detection region, it fires.

That algorithm has no concept of an object. It cannot distinguish a person from a raindrop, only a large brightness change from a small one. Everything below is a way of producing a large brightness change without a person being present, and every threshold you can set is a compromise between missing real events and reporting fake ones.

Infrared bloom, insects and spider webs

This is the number one cause of a camera that alerts all night and nothing during the day.

After dark the camera switches to infrared and lights the scene with an emitter, usually mounted in the same housing as the lens. Brightness falls off sharply with distance, so anything within a few centimetres of the lens is lit far, far brighter than anything at distance. A moth at the lens is a white shape covering a third of the frame. A spider that has built a web across the housing is a permanent bright object that moves in every breath of wind, all night, every night, until the web is cleaned off.

Fixes, in order of effectiveness. Move the illuminator away from the lens, which means using a separate infrared lamp mounted a metre or two away rather than the one built into the camera body. Clean the dome and the housing, and keep cleaning it. Raise the minimum object size so a close insect no longer qualifies, accepting that this reduces your range. And classify, because a model asked what class of object this is does not answer person for a moth.

Rain, snow and fog

Every raindrop and snowflake in the infrared beam is a small bright moving object, and there are thousands of them. Fog is worse in a subtler way: it scatters the infrared back into the lens and lifts the brightness of the whole frame, so the detector sees a large change and the image loses the contrast it needed to see anything real.

Fixes. A minimum duration of one to two seconds removes most precipitation, because an individual drop crosses the frame in a fraction of a second while a person is present for several. Housing shape helps: a camera under an eave or with a decent sun shield catches less. Beyond that, classification is the only thing that reliably separates heavy rain from a person walking through it, which is exactly the condition in which you most want to know.

Headlights, reflections and passing vehicles

A car on a road fifty metres away never enters your property, but its headlights sweep across your wall, your gate and the wet tarmac of your forecourt. What moves in frame is a large, bright, fast patch of light, and that is a textbook large brightness change.

Reflective surfaces multiply this. Corrugated steel, wet asphalt, glass doors and parked car bodywork all take a light source that is outside the frame and put it inside the frame as a moving object.

Fixes. A polygon exclude zone over the road and the verge, drawn at the kerb rather than at your fence so somebody approaching your gate still registers. A vehicle class filter, so vehicles in the timeline do not have to be vehicles on your phone. And where possible, a camera angle that does not include the road at all, which is free and permanent.

Auto exposure, auto white balance and IR cut switching

When the camera re rates the scene, every pixel changes at once. Automatic exposure adjusting to a cloud, automatic white balance shifting at dusk, and the infrared cut filter snapping in or out at dawn and dusk each produce a whole frame change that a naive detector reports as a very large event.

The signature is unmistakable: an alert at almost exactly the same two times every day, sliding by a few minutes a week as the seasons move.

Fixes. Lock the exposure and the white balance if the camera allows it. Set the day and night switching to fixed times rather than by light level, so it happens once instead of repeatedly on a changeable evening. And use a detector with an adaptive background model, which absorbs a whole frame brightness shift instead of reading it as motion. That is how the detection stage here handles it.

Shadows, clouds and sunbeams

A cloud crossing the sun changes the brightness of an entire yard in a couple of seconds. A gate casting a shadow that creeps across a drive is a slow moving dark object. A sunbeam through a window lands on a stockroom floor at the same time every afternoon and moves steadily for an hour.

Fixes. An adaptive background handles slow shadows well, because they change slowly compared to a person. Fast cloud shadows are best handled by classification. Indoor sunbeams are best handled by excluding that patch of floor, since nothing important happens there anyway.

Vegetation, wind and everything else that moves

Branches, long grass, flags, banners, hanging signs, plastic sheeting, roller doors flexing, and pallet wrap on a windy dock. All genuinely moving, all genuinely uninteresting.

The seasonal trap is worth knowing. A camera tuned in February looks at bare branches. In June the same view is a wall of leaves and the same settings produce a hundred alerts a night, and it feels like the camera degraded when in fact the scene changed.

Fixes. Exclude zones over the vegetation, minimum duration for the gusts, and classification for everything else. Trimming the branch is not a joke, it is frequently the cheapest permanent fix available.

Arlo camera false triggers, and the rest of the battery cameras

Everything above describes a camera that compares frames. A battery powered consumer camera usually does not start there. Arlo, Ring, Blink and most of their competitors wake on a passive infrared sensor, which does not see pixels at all: it sees moving heat. That changes which false triggers you get, so the fixes above only half apply.

Passive infrared fires on a moving heat difference against the background, which is why these cameras behave the way they do. Sun warmed brick, a tarmac drive giving back the day's heat, a car engine or its headlights, warm air rolling out of a vent, and rain running down the lens cover all read as moving heat. The counterintuitive one is summer: when the air is close to body temperature the contrast between a person and the background collapses, so the same camera misses real people in a heatwave and trips on the wall behind them.

Fixes. Lower the sensitivity rather than raising it, because with passive infrared higher sensitivity mostly buys you the driveway warming up. Angle the camera across the walking path instead of along it, since these sensors detect movement crossing their zones far better than movement coming straight at them. Keep it out of direct sun and off a surface that stores heat. Draw the activity zone the vendor gives you. Then, if the vendor sells object detection on a subscription, that filter is the only thing that will separate a person from a warm wall.

The honest limit, and it applies to us: most of these cameras are cloud locked and never expose an ONVIF or RTSP stream, so no third party software, ours included, can read them. If your cameras are Arlo or Ring, the answer is their own settings and their own subscription, not this product. We can only help once a camera hands out a stream URL, which is what RTSP software and every business IP camera do. That is a real reason to choose an ONVIF camera the next time you buy one.

The order to fix them in

1. Clean the lens and the housing. Free, takes five minutes, and it is the cause more often than anybody expects.
2. Fix the framing. Get the road, the sky and the neighbour out of the picture. Permanent and free.
3. Move the infrared source away from the lens. The one hardware change worth making before you buy any software.
4. Lock exposure and white balance, and set day and night switching to fixed times.
5. Set a minimum duration of one to two seconds. The highest value setting in this whole list.
6. Draw exclude zones over what is not yours. Polygons if your software has them, rectangles if not.
7. Then, and only then, add classification. Because it fixes the causes the first six cannot.

The ceiling on tuning, and what is above it

Work through those seven and most sites get quiet nights. What is left over is the set of causes that no threshold can separate from a real event, because they look identical to a pixel counter: a moth against a person in dark clothing at distance, headlights against a torch, rain against somebody walking through rain.

Separating those needs a detector that answers a different question. Not how much changed, but what was it. That is the whole of the difference between motion detection software that counts pixels and artificial intelligence for video surveillance that classifies objects, and it is why every event in our timeline carries a class and a confidence value rather than a percentage of pixels changed.

You can test that difference on your own worst camera in about a minute. Export sixty seconds from the hour that generates the junk, drop it into the demo, and read what comes back. The detection runs in your browser, on your machine, and the clip does not leave it.