The Off-Key Bard
Technology

Why Do Security Cameras Sometimes Detect Motion That Isn't There?

Or: why your security camera thinks a spiderweb, passing headlights or moving shadow is an intruder. Cameras don't see objects moving. They detect changing pixels, shifting heat or patterns a model guesses are people, and each method has its own ways of being fooled.

Or: Why Your Security Camera Thinks a Spiderweb, Passing Headlights, or Moving Shadow Is an Intruder

It's two o'clock in the morning and your phone lights up. Motion detected: front driveway. You open the app expecting someone at the house. The clip shows an empty driveway, a parked car and the same trees that have stood there for years. A few minutes later, another alert. By morning, the camera has logged seventeen motion events without capturing a single person.

Or maybe it pings you every time a branch sways yet misses the delivery driver walking straight up to your door, or goes haywire after sunset. It looks like a malfunction, but usually it's a misunderstanding of what the camera is actually detecting. A security camera doesn't necessarily see a person or object moving. Depending on its design, it may detect changes in an image, changes in infrared radiation, or patterns that software interprets as meaningful. Those are very different things, and knowing which one your camera uses explains why it reports motion when nothing obvious moved. It also explains why the best fix is often changing how the camera detects activity rather than replacing the camera.

Motion Detection Isn't One Technology

Manufacturers who advertise motion detection aren't all describing the same system. A wired surveillance camera may analyze differences between video frames. A battery camera may rely on a passive infrared sensor to decide when to wake its main imaging system. A more advanced camera may use machine learning to identify people, vehicles, animals or packages. Some combine methods: an infrared sensor wakes the device, software checks whether a person is present, and a notification goes out only if the event meets your settings.

Picture a pipeline. Something changes in the environment: a shadow, headlights, wind, an insect, rain. A detection mechanism, either video frame differencing or a thermal sensor, registers it. A software or AI filter may estimate whether it matches a person or vehicle. Finally, sensitivity thresholds decide whether to save a clip or send an alert. Every stage can err for a different reason, so a problem on a battery Blink camera may need a different fix from the same problem on a PoE camera feeding an NVR.

Video Detection Sees Changing Pixels, Not Objects

The simplest approach compares the current scene against a model of the background. OpenCV's documentation describes the standard technique, background subtraction: subtract a background model, containing the static part of the scene, from the current frame to produce a foreground mask, a binary image of the pixels belonging to moving objects. The model is computed initially and then updated continuously to adapt to changes in the scene, and the method assumes a static camera.

This works because the software doesn't need to understand what moved. It only needs to notice that pixels changed, and that's also the source of most false alarms. A bright patch of sunlight crawling across the driveway as clouds pass changes thousands of pixels although nothing entered the scene. So do shifting shadows, sweeping headlights and automatic lights switching on. The camera can be working exactly as designed while sending alerts that seem pointless to you. Better detectors add filtering and object tracking to separate meaningful motion from noise, but the software is still interpreting visual information, not measuring an intruder's presence. Google's help page for Nest cameras says as much: the camera relies on software looking for changes in color and lighting, so triggers can include light-level changes, moving shadows, swaying trees or bushes, and headlights or flashes entering the view.

Shadows, Headlights and Wind

A shadow seems easy to ignore, since it isn't an object. But when a big tree shades your driveway, the branches move and the shadow's shape changes: dark regions turn bright and bright ones turn dark. To a basic detector, that can look like a large object passing through. Some algorithms use color, shape and behavior over time to tell shadows from objects, but difficult lighting still fools them.

Headlights are similar. A car driving past never enters your driveway, but its beams sweep across the garage door and lawn, changing brightness across large parts of the frame, and at night the camera's exposure system reacts too, altering the whole scene. The trigger may be illumination from something outside the field of view. The same happens when a floodlight switches on. As yesterday's article explained, many cameras also shift between color daytime operation and infrared night mode, changing contrast and brightness, and one near a streetlight can flip modes repeatedly as light hovers around its threshold.

Wind causes real visual movement. The camera isn't wrong that something moved. It's wrong about whether the movement deserves your attention. If a tree fills half the frame, even a breeze generates a lot of changing image data, and raising the threshold can make the camera less responsive to genuine activity elsewhere. A better fix is usually restricting detection to the areas that matter.

How Passive Infrared Sensors Work

Many battery-powered security cameras use passive infrared sensors, or PIR, which operate differently from video-based motion detection. A PIR sensor does not emit infrared radiation or produce an image. Instead, it responds to changes in thermal infrared energy entering its field of view. The sensing element uses a pyroelectric material that generates an electrical response when its temperature changes. An infrared-transmitting lens, often a Fresnel lens, divides the detection area into multiple zones, allowing a warm or cool object moving across those zones to produce a changing signal. PIR sensors commonly respond to thermal radiation in a range that includes approximately 8 to 14 micrometers, although the exact wavelength characteristics depend on the sensor. A stationary heat source generally produces little response once conditions stabilize, but moving animals, rapidly changing heat sources, sunlight heating surfaces, or changes in the surrounding thermal environment can still cause unwanted triggers. The sensor detects changing infrared energy, not the identity of whatever caused it.

PIR isn't a thermal camera. A conventional PIR sensor produces a simple signal representing changes in infrared energy across its zones, not a temperature map, and it can't tell a human from an animal. Its virtue is detecting potential movement on very little power, but it isn't intelligent person recognition.

The zone structure suggests some practical behavior, though this is my reasoning from the design rather than something the sources state. Walking across the field of view crosses several zones in quick succession and produces a pronounced change, while approaching head-on may produce a weaker one, which is why manufacturers often advise mounting PIR cameras so people cross the detection area. Temperature contrast should matter too: a person is easier to detect against a cool background than against sun-warmed pavement.

Missed Beginnings and Empty Clips

A battery camera saves energy by keeping its main sensor and processor in a low-power state until a detector fires. The PIR sensor notices a delivery driver, the camera wakes, and only then does the clip begin, so some of the initial movement may already be over. A powered camera that continuously processes video can keep a rolling pre-event buffer, but a deeply sleeping battery camera has less ability to. That's why a camera can record someone leaving the door but miss their arrival.

The reverse also happens: an alert arrives, and the clip shows nothing, because the object left view before recording started or an insect was near the lens for a moment. Google's help page adds that detection can fail when activity is outside the field of view, blocked, far away or in dim light, and that cameras intentionally don't re-alert during continuing activity. A notification means some part of the detection process hit its trigger conditions, not that the clip shows an identifiable subject.

Insects, Spiderwebs and Weather

Outdoor cameras attract insects, especially around illumination. Perspective does the rest: for a simplified pinhole camera, apparent size is proportional to physical size divided by distance, so a moth an inch from the lens can occupy more pixels than a person at the far end of the driveway. Strongly lit by infrared and out of focus, it becomes a huge glowing shape darting across the frame. Detectors don't automatically know the true size of what they see.

Spiderwebs are worse. A web nearly invisible by day lights up under infrared at night, wind moves it slightly, and because it's right at the lens the changing shapes fill much of the image, triggering event after event. It often appears suddenly after a spider moves in near the housing. The first step isn't lowering sensitivity. It's gently cleaning the camera as the manufacturer advises, which removes the source.

Rain and snow add real moving streaks that illumination makes visible at night, and water on the lens distorts the scene. Ring's documentation also warns that moving trees, sunlight, reflections and shadows can cause unwanted alerts, and that a device near a window or mirror may be affected by reflections. Sometimes the camera itself moves: a lightweight bracket vibrating in the wind shifts the whole image, so buildings, fences and trees all seem to move, and background subtraction assumes a static camera. A stable mount is part of reliable detection.

Sensitivity, Zones and Why They Differ

Sensitivity controls don't have one universal meaning. On a video system, they may set how much image change triggers an event, on a PIR camera the threshold for interpreting sensor activity, and on an intelligent system separate settings may govern motion and object classification. Higher sensitivity catches smaller activity but increases false alerts, and lower sensitivity risks missing real events.

Detection zones are often more valuable than changing sensitivity alone. Suppose a wide-angle camera sees your driveway, a public street and a large tree. Lowering sensitivity until the street and tree stop triggering might also cause the camera to miss someone walking near your car. A better approach is to define a zone covering the driveway and walkway while excluding areas that generate unwanted activity. Ring's documentation describes motion zones as a way to reduce unnecessary alerts from passing vehicles, shadows and other activity, with support for up to three zones on applicable devices. Google Nest offers similar activity-zone controls, but an important distinction applies: zones do not necessarily prevent a camera from detecting or recording everything outside them. Depending on the model and recording configuration, they may primarily control which events generate notifications. Before relying on a zone, check whether it affects motion detection, event recording, alerts or some combination of those functions.

One caveat matters for battery cameras. Ring notes that battery-powered devices may need Advanced Motion Detection turned on to use motion zones, and its standard maximum detection range is 30 feet. A video-based zone defines which pixels the software analyzes, but a separate PIR sensor has its own physical detection pattern, so drawing a rectangle in the app doesn't necessarily change what the PIR sees. On some devices, movement outside the software zone may still wake the camera. That's general reasoning about the architecture, so check how your model behaves.

AI Detection: Classification, Not Understanding

Modern cameras increasingly advertise AI detection: software trained to recognize visual patterns associated with people, vehicles, animals or packages. Instead of alerting whenever pixels change, it tries to determine what caused the change, which can dramatically cut notifications by ignoring swaying branches and flagging the person on the driveway.

But it's classification, not understanding. The system estimates whether observed patterns resemble the classes it was built to recognize. Google's help page is candid: the technology isn't perfect, a picture can be confused with a real event (its example is a t-shirt with a dog on it triggering an animal alert), and a person crawling on the ground can be misidentified as a pet. A classification is an interpretation, not proof.

Confidence thresholds create a tradeoff. An object-detection model may assign a confidence score to a possible person, vehicle or animal, and the system uses that score along with its configured thresholds and other rules to decide whether an alert should be generated. Set low, the camera catches people in difficult conditions but produces more false positives, like a shadow flagged as a person. Set high, it ignores uncertain objects, cutting alerts but risking false negatives, real events that go undetected. False positives cause notification fatigue, while false negatives leave important events unnoticed, and reducing one often raises the other. Tuning purely to eliminate every unwanted notification is a mistake.

Conditions matter. At night, the problems from yesterday's article spill into detection: long exposures blur a person, infrared adds odd contrast, and noise reduction and compression strip detail, so the classifier gets worse input. Better lighting can improve intelligent detection without touching any AI setting. Distance matters too: someone near the street under a wide-angle lens may be too small to classify, so evaluate person detection at the distance you actually need. Some systems track objects across frames, which rejects brief headlight flickers but may delay detection of someone who's only briefly in view. Advanced cameras and NVRs also offer line-crossing detection and intrusion zones, which fire only when something crosses a virtual line or enters a defined area like a gate, though a poorly placed line still misfires.

Alerts, Recording and Alarm Fatigue

Notifications and recording needn't follow the same rules. A wired NVR might record continuously while notifying only when it detects a person. That separation lets you preserve evidence without being pestered every time a squirrel crosses the lawn, and it shows why continuous recording is valuable when detection fails: motion-triggered recording only works if the detector recognizes the event, whereas continuous recording stores everything and lets detection act as an index.

False alarms are more than a nuisance. Imagine twenty pointless alerts a night. You check each at first, then ignore them, then disable notifications, and then a legitimate event arrives that you've stopped watching for. This is alarm fatigue, and it's why tuning is about preserving the usefulness of the alerts, not just reducing irritation.

How to Fix It

Start by finding the pattern. Do alerts happen only at night, only in wind, only in heavy rain, or when cars pass? Do the clips show bright streaks or large out-of-focus shapes? Glowing shapes at night point to insects, webs or infrared reflections. Alerts matching traffic suggest headlights or activity outside the intended area. Daytime alerts in wind suggest foliage or camera vibration. Then change one factor at a time, rather than lowering every sensitivity setting and hoping.

A sensible order: check the mounting and field of view first and remove physical causes like webs or touching vegetation. Position is often the best fix, since a slight rotation can drop a road from the frame and moving a camera away from a tree can end recurring branch alerts. Then set detection zones to exclude areas you don't need, and use person or vehicle categories to control which events notify you. Adjust sensitivity in small steps, testing after each change by having someone walk through at realistic speeds and distances, including at night, because a setup tuned in daylight can behave very differently after dark. Some premium systems add radar, which detects movement from reflected radio waves instead of image pixels, but for ordinary homes well-configured video or PIR detection is plenty.

The Bard's Take

Security cameras have become remarkably good at recognizing activity, but those features are built on technologies that interpret imperfect information. Video detection observes changes in an image, so shadows, headlights and swaying branches can alter it when nothing has entered the property. PIR responds to changes in infrared radiation, which suits low-power battery cameras but depends on movement direction, temperature contrast and environment. AI adds classification that reduces false alerts substantially, but its decisions depend on image quality and learned patterns. At night it all gets harder.

The camera may be working normally while its detector reaches the wrong conclusion, so fixing false alarms means understanding what it's actually detecting. If foliage is the culprit, adjust the field of view or zone. If insects trigger recordings, clean the housing. If passing cars cause alerts, exclude the street. If a battery camera keeps missing people approaching head-on, look at the PIR orientation. And if you must never miss something, consider continuous recording as the foundation rather than relying only on triggered clips.

The best system isn't the one that detects the most movement. It's the one that reliably separates meaningful activity from ordinary environmental change while preserving the footage you may need later. A silent system that misses the important event is useless, and one that alerts every thirty seconds becomes background noise.

A security camera doesn't need to detect every change in its surroundings. It needs to recognize the changes that matter, and help you ignore the ones that don't.

Sources

← Back to the Desk — Technology