Person detection is a video analytics function that uses a trained machine learning model to decide whether a human figure is present in a camera image, and raises an event only when it finds one. In home security, it turns a security camera or video doorbell from a device that reacts to any change in pixels into one that reacts to people specifically.
The feature exists because simple motion detection produces large numbers of false alerts from swaying trees, headlights, rain, insects and pets. By filtering events through a person classifier, smart home systems send fewer and more meaningful notifications, record more relevant clips and can trigger automations such as lights or sirens with much greater confidence.
How person detection analyzes video
Person detection usually runs as a two-stage pipeline. A lightweight first stage, either pixel-difference motion analysis or a passive infrared (PIR) sensor, wakes the system when something changes. The second stage passes selected frames to an object detection neural network, typically a compact convolutional or transformer model trained on large labeled image sets, which outputs bounding boxes, class labels and a confidence score.
The system compares that score with a confidence threshold, often around 0.5 to 0.7, and checks additional rules such as the activity zone, the minimum object size and how many consecutive frames contain the figure. Only when all conditions are met does it create a person event, attach a snapshot and start or tag a recording.
The following table summarizes the main stages and what each contributes.
| Stage | Function | Effect on results |
|---|---|---|
| Trigger | PIR or pixel motion wakes the analysis | Saves power, may miss slow movement |
| Inference | Neural network classifies objects in frames | Separates people from other motion |
| Filtering | Zones, size and threshold rules | Removes irrelevant detections |
| Tracking | Links detections across frames | Prevents duplicate alerts for one visit |
| Notification | Sends alert with snapshot or clip | Determines what the owner sees |
On-device and cloud processing
The neural network can run in three places, and the choice affects latency, privacy, cost and battery life. Modern cameras increasingly include a small neural processing unit (NPU) that performs inference on the device, while older or cheaper models upload frames or clips to a cloud service for analysis.
| Processing location | Advantages | Limitations |
|---|---|---|
| On the camera | Fast alerts, works offline, video stays local | Smaller models, limited accuracy at distance |
| Local hub or NVR | Larger models, shared across cameras | Extra hardware, network load |
| Cloud service | Most capable models, easy updates | Subscription, upload delay, privacy exposure |
On-device processing is the strongest option for privacy, because frames never leave the home unless a clip is deliberately saved. Cloud processing typically adds one to several seconds of delay and depends on internet availability, which matters when the alert should trigger an immediate response.
Person detection compared with other detection types
Person detection is one of several smart detection categories that cameras offer. It answers only the question of whether a human is present, and it should not be confused with features that identify who the person is or what they are doing.
| Feature | What it recognizes | Typical use |
|---|---|---|
| Motion detection | Any change in pixels or heat | Basic triggering, high false-alarm rate |
| Person detection | A human figure of any identity | Intrusion alerts, doorstep events |
| Vehicle and animal detection | Cars, pets, wildlife | Driveway monitoring, pet filtering |
| Package detection | Parcels placed or removed | Delivery notifications |
| Facial recognition | Identity of a known face | Familiar-face alerts, legally sensitive |
The distinction from facial recognition has legal weight. Detecting that a person is present generally involves no biometric identification, whereas matching faces against a stored library processes biometric data and falls under stricter rules in several jurisdictions, including the GDPR in the European Union and some state laws in the United States.
Accuracy, errors and tuning
No model is perfect, and the two error types pull in opposite directions. A false positive alerts on a coat on a chair or a statue, while a false negative misses a real visitor. Raising the threshold reduces false alarms but increases the risk of missed people, so tuning is always a trade-off between nuisance and coverage.
Common causes of errors and their remedies are summarized below.
| Symptom | Likely cause | Fix |
|---|---|---|
| Alerts from posters or mannequins | Static human-like shapes | Exclude area with a privacy or detection zone |
| Missed people at night | Low contrast in infrared image | Add lighting or adjust camera angle |
| Missed people far away | Figure too small in frame | Mount closer or use higher resolution |
| Alerts from passers-by on the street | Zone covers public sidewalk | Redraw activity zone to the property |
| Late or no alerts on battery cameras | PIR trigger missed slow approach | Raise PIR sensitivity or use wired power |
Mounting height and angle have a major influence. A camera placed at about 2 to 3 meters and tilted slightly downward shows the full body profile that models are trained on, while extreme top-down views or heavy wide-angle distortion reduce recognition rates.
Automations and privacy considerations
Because person events are far more reliable than raw motion, they are a better trigger for smart home automations. Integrated with a hub, an alarm system or a voice assistant, a person event can drive actions that would be too disruptive if triggered by every leaf movement.
- Turn on outdoor lights when a person enters the driveway after sunset.
- Announce a visitor on indoor speakers before the doorbell is pressed.
- Escalate to a siren or professional monitoring only when the alarm is armed.
- Record longer clips for person events while discarding plain motion.
Responsible use also means respecting neighbors and passers-by. Activity zones and privacy masks should exclude public areas and adjacent properties where possible, and owners should check local rules on recording beyond their own boundary, because lawful recording limits vary between countries.
The simple explanation
Person detection is a camera skill that asks one question: is there a human in this picture? An ordinary motion alarm wakes up for anything that moves, such as a cat, a car’s headlights or branches in the wind.
A camera with person detection looks more carefully, much like a doorman who ignores the wind but notices when someone walks up the path. The result is fewer pointless alerts, so when the phone does buzz, it is much more likely to matter.
How it is used in practice
A homeowner with a front-yard camera facing a tree-lined street received more than 60 motion alerts a day during windy autumn weeks, mostly from branches, passing cars and a neighbor’s cat. The notifications became so frequent that she started ignoring them.
She enabled on-device person detection, redrew the activity zone to cover only the path and porch, and set the camera to alert on person events only. She also linked person events after 10 p.m. to the porch lights through her smart home hub.
Alerts fell to about 5 to 8 per day, nearly all deliveries and visitors. One night the porch light switched on as someone approached the side gate, the clip was saved automatically, and the person left before reaching the house, which showed that fewer but trustworthy alerts are what make a home security camera useful.