If, a week after installation, the security team mutes the alarm notifications, the person detection project is effectively over. The cause is usually not the software: the camera is at the wrong height, the target is too small in the image, the night IR reflects the rain, and the threshold was left at the manufacturer default. AI-powered analytics distinguishes whether an object in the image is a person, a vehicle or a tree branch; but how well it does so varies by site. Reliability isn’t a product feature; it’s a measured result.
The difference between motion detection and object classification
Classic motion detection measures pixel change between consecutive frames. A tree swaying in the wind, a raindrop, a shadow, headlights or an insect passing in front of the camera also produce pixel change. Deep learning-based analytics finds and classifies objects in the image and runs rules only for specific classes. This approach can noticeably reduce false alarms. It doesn’t eliminate them.
A point often confused on site: classification and rules are separate layers. The model says “person”; the rule answers “did this person cross the virtual line in this direction, and how many seconds did they stay in this zone?”. A large share of false alarms comes not from the model but from a badly drawn zone or an overly sensitive dwell time. Playing with the threshold without separating the problem just swaps one error for another.
Factors that determine accuracy
The target’s pixel size
A model can reliably classify a person only if it sees them at a certain minimum size in the image. Manufacturers give different minimum values depending on the model and version. Distant targets, wide-angle lenses and the use of a low-resolution substream make the target smaller.
The substream issue is often overlooked. On some cameras, edge analytics runs on a low-resolution internal stream instead of the main stream; in server-based systems, the substream is also chosen to reduce license or GPU load. A camera that is 4 MP on paper may be processing a 640-pixel-wide image for analytics. During installation, check which stream the analytics uses and measure the target’s size at the farthest detection point on that stream.
Camera angle and mounting
Cameras looking straight down from very high show a person’s silhouette from above and make classification harder. A very horizontal angle makes distant targets smaller and causes objects passing in front to hide those behind. A tilt of about 15–35 degrees is generally a good starting point for analytics; the exact value is set by the scene.
Light, night and weather
Under IR, the image is black and white and noise increases; raindrops reflect IR light, and spider webs glow in front of the lens. Backlight and headlight glare distort the silhouette. Analytics performance should be measured separately for day and night.
The thing that most degrades night performance is built-in IR. Because the IR source is close to the lens, every droplet and insect in the air appears as a bright spot. Positioning an external IR illuminator a little away from the camera, or lighting the area with white light, can give more improvement on site than threshold tuning.
Scene complexity
Crowds, objects hiding each other, reflective floors, screens and posters (posters showing people) can lead to misclassification. Reflections in a glass façade have the same effect: the reflection of a person outside registers as “person in zone” on the camera inside.
Measure with the right metrics
| Metric | Definition | Why does it matter? |
|---|---|---|
| Precision | The share of generated alarms that are real events | Operator trust in the system |
| Recall | The share of test events that generated an alarm | Security risk |
| False alarms / camera / day | Daily number of alarms that aren’t real | Operational sustainability |
| Latency | Time between the event and the notification | Response time |
Precision and recall should be assessed together: lowering the threshold increases recall but also increases false alarms. The right balance depends on the scenario’s risk; in perimeter intrusion a missed event is more costly, while in a busy warehouse false alarms quickly wear out the operator.
Example calculation: during the pilot, 40 test crossings were made and the system raised an alarm for 36 of them; in the same period, of 50 total alarms, 36 were real and 14 false. Recall is 36/40 = 90% and precision is 36/50 = 72%. If the same result is reported as “90% successful”, it hides that one in four alarms the operator sees is empty. Write down both numbers.
How do you design a pilot?
- Scenario selection: 4–10 representative cameras covering different angles, lighting and scene types.
- Writing down acceptance criteria: Target precision/recall for each scenario, or an upper limit on daily false alarms per camera.
- Test events: Staged crossings by day and night, in different directions and at different speeds; a rainy night is included in the measurement if possible.
- Logging and classification: Marking every alarm as real/false and grouping false alarms by root cause.
- Tuning and re-measurement: A second measurement after zones, thresholds, size filters and schedules are adjusted.
The weakest link of a pilot is the test events. If everyone walks straight towards the camera, upright, the system looks good. A real intrusion doesn’t happen like that: people crouch, run, walk parallel along the fence or pass close to the camera’s blind corner. Add these movements and the farthest detection point to the test scenarios. How to fix the root causes of false alarms is explained in detail in the false alarm guide.
Edge or server?
Edge analytics runs on the camera; it reduces network and server load but is limited by the camera’s processing power. Server-based analytics can run larger models using GPUs and can be added to existing cameras; however, per-channel processing capacity and license costs must be calculated. In a hybrid approach, the camera does the first classification and the server handles search and advanced rules. For bringing events into the VMS, ONVIF Profile M support matters.
When deciding, ask these questions: Will the existing cameras be replaced? Should alarms also be generated when the network connection is lost? Is analytics event metadata searchable in the VMS, or do only notifications arrive? Who will update models and firmware, and after what testing? A firmware update can change model behavior. A short re-measurement after an update catches false alarms quietly creeping up on site early.
Ethical and legal limits
Person detection doesn’t establish identity. Applications such as facial recognition or individual tracking of employees require a separate assessment in terms of biometric data and employment law, and shouldn’t be set up as a default service. Frameworks such as the NIST AI Risk Management Framework recommend that AI systems be limited to their purpose, measurable and traceable.
In practice, this means tying every analytics rule to a purpose statement, such as “detect after-hours warehouse entry”. A rule without a purpose statement turns, over time, into a system where it’s unclear who is monitoring what.
Frequently asked questions
The manufacturer says 99% accuracy; why is it lower on site?
Manufacturer figures are measured on specific datasets under ideal conditions. Since the angle, light, distance, weather and scene complexity on site are different, the result changes. That’s why metrics measured on your own site should be the basis.
Is edge or server-based analytics more accurate?
Neither is more accurate on its own; it depends on the model, processing power and image quality. Server-based systems can run larger models; edge analytics reduces network load. They should be compared with a pilot.
Does person detection identify the person?
No. Person detection is the classification “this is a person”; it doesn’t determine who it is. Biometric recognition is a separate technology and requires a separate legal assessment.
To start analytics at your facility by measuring first, see our AI video analytics pilot service.