Video analytics projects usually fail operationally, not technically. Through the night, the operator closes dozens of alarms for swaying branches, headlights and insects drawn to the IR; after a while they mute notifications or disable the rule. When a real intrusion comes, nobody is looking at the screen. AI-based classification can noticeably reduce false alarms but doesn’t solve the problem on its own. This guide explains how to classify false alarms by root cause and reduce them by measurement.
Measure first
Improvement needs a baseline. When each alarm is closed, it’s marked as real, false or unclear. With one to two weeks of data, the daily false alarm count and precision per camera are worked out. Which cameras produce most of the total false alarms usually becomes clear at this stage; in most systems, a small number of cameras account for most of the problem.
Setting changes made without measurement are guesses. “Fewer alarms came in last night” could just be because the wind dropped. When logging, keeping the alarm time, camera, rule and a short cause note (“headlights”, “cat”, “rain”) turns root cause analysis into a few minutes of filtering. Keep night and day data separate. Their problems are different.
Root causes of false alarms
| Root cause | Example | Remediation |
|---|---|---|
| Environmental movement | Trees, flags, plants, water reflections | Narrowing the zone, class filter, pruning |
| Light changes | Headlights, cloud shadows, automatic lighting | Camera angle, WDR, night profile |
| IR and weather | Rain, snow, insects, spider webs | External IR, lens cleaning, size filter |
| Misclassification | Shadows mistaken for people, posters | Minimum size, duration filter, model update |
| Rule design | A virtual fence crossing a road, an overly wide zone | Double lines, direction filter, redrawing the zone |
| Schedule | A perimeter rule active during working hours | Time-based rule sets |
| Expected activity | Security patrols, maintenance team | Procedure, logged temporary disabling |
A common picture is several causes stacked on the same camera. For example, on a camera facing the perimeter wall, insects drawn to the IR light at night create large, bright spots in front of the lens, while at the same time the headlights of passing vehicles suddenly light up the scene. Trying to solve both with a single setting loosens the rule so much that the real target is missed too. Separate the causes and solve them one by one.
Camera and scene fixes
- Angle: An angle perpendicular to the target’s direction of movement that shows the target at sufficient pixel size improves classification (person detection guide).
- IR placement: IR built into the camera makes insects and raindrops in front of the lens glow; an external IR projector can reduce this effect.
- Lighting: Steady, sufficient lighting improves night performance; motion-sensor lamps can trigger alarms through sudden light changes.
- Cleaning: Dirt on the lens and housing is a frequent cause of false alarms, especially at night.
- Vegetation and obstructions: Branches entering the zone are pruned, or the zone is narrowed.
Analytics rules are seen as a feature “added” to the camera after installation. In fact, much of the classification model’s job is determined by where the camera is placed. At an angle where the target walks towards the camera, the human figure stays small and shapeless for a few seconds; when the same target passes sideways, the silhouette is clear. If the target drops to a few pixels high at the farthest point of the site, the model can’t confidently call it “person”. In that case, removing the rule from the far zone or covering that area with a second camera gives better results than fiddling with the threshold.
Spider webs deserve their own heading. Insects drawn by IR bring spiders. A single strand of web swaying in the wind in front of the lens produces motion detections all night. It comes back within a few weeks of cleaning. Separating the IR from the camera with an external illuminator, or increasing the maintenance frequency for that camera, is closer to a permanent fix.
Rule fixes
- Class filter: Let the rule run only for people or vehicles.
- Zone: Limit it to the area that really needs protecting; leave roads and pavements outside.
- Direction: Let only inward crossings raise an alarm.
- Duration: A condition to stay in the zone longer than a set time filters out momentary crossings.
- Double line: Requiring the target to cross two lines in sequence is more reliable than a single line.
- Size: Minimum and maximum object size according to perspective.
- Verification: Requiring thermal + visual cameras, or two cameras, to see the same event.
The classic mistake in rule design is drawing the zone over the whole image “so it catches everything”. If the edge of the zone touches a road, pavement or neighboring plot, every pedestrian and vehicle passing there falls into the rule. Draw the virtual fence just inside the wall, on a line the target must cross. For the size filter, take perspective into account: a cat close to the camera looks bigger than a person far away. Minimum and maximum size should be assessed separately for the front and back of the scene; on supporting systems, do this with perspective calibration.
Setting the duration filter too long is also a risk. A target running through stays in the zone only briefly. Set the filter duration by testing against the real crossing time in the scene.
Process fixes
Some alarms are technically correct but operationally expected activity: security patrols, the night maintenance team, deliveries. For these, a procedure is defined (notification, schedule exception). An alarm verification flow lets the operator see the event clip quickly (video verification).
The schedule is the most effective tool in this step. There’s no point keeping a perimeter intrusion rule active during working hours, when staff walk around the grounds. Define a separate rule set for weekends and holidays. Temporary disabling must always be logged and automatically re-enabled; there are plenty of sites where a rule “turned off for maintenance” stayed off for weeks.
Don’t forget missed events
While reducing false alarms, the risk of missing real events can increase. After every setting change, recall should be checked with test events (different directions, speeds, day/night). Changes and their results are logged, so you can trace which setting changed what.
A simple list of test scenarios is enough: walking through, running through, crouching forward, passing at the edge of the zone, a night crossing in dark clothing. Keep each scenario in writing and repeat the same list after every tuning round. Zero alarms isn’t success. The rule might not be catching anything.
The loop: measure, fix, measure again
- Two weeks of baseline data
- Root cause analysis on the cameras producing the most alarms
- Camera, scene and rule fixes
- Checking recall with test events
- Two weeks of new measurement and comparison
- Repeat at seasonal changes (leaves, snow, daylight)
Don’t skip the seasonal step. A scene tuned in summer behaves differently with falling leaves in autumn, and snow and early darkness in winter. When daylight hours change, the time of switching to the night profile shifts too. Running this loop again at least twice a year is the practical way to keep the system performing as it did on day one.
Frequently asked questions
Is it possible to eliminate false alarms completely?
No. The goal is to bring them down to a level the operator can sustain while keeping the risk of missed events acceptable. This balance is set according to the scenario.
Isn’t raising the threshold enough?
Raising the threshold reduces false alarms but increases the risk of missing real events. Fixing the root cause first (angle, light, zone) is sounder.
Who should keep false alarm statistics?
The best data source is operators marking alarms as real/false/unclear when closing them. VMS alarm management or a simple log form can be used.
We run this loop in the pilot and operations phases as part of our AI video analytics and virtual fence services.