Machine learning explained: what are true and false positives?

In video analytics and remote monitoring, we talk a lot about true positives and false positives, and in turn you often hear people regularly mention true and false alarms. So what exactly do these terms mean, in a technical sense? In this article, we’re going to break down a key concept that underpins how our False Alarm Filtering Platform works. 

What are true and false alarms?

First, let’s look at what we mean by true and false alarms. We tend to set up security cameras in sites where people aren’t supposed to be. These cameras have detectors that sense movement, so that we can find out when people are trespassing on the site. When the sensor is triggered, it captures around 2-3 images of the movement it’s spotted, and sends these images through as an alarm. In most cases, these alarms are sent to dedicated monitoring stations. 

A true alarm contains movement that was caused by human activity - either people or cars. These are the alarms operators need to look out for and potentially respond to. However in most cases, security cameras send false alarms. These are caused by non-human activity - light reflections, trees or tarpaulin blowing in the wind, and animals running across the site. 

Over 95% of the alarms security cameras send through are false, which makes a monitoring operator’s job much more challenging than it needs to be. Video analytics platforms like Calipsa’s use artificial intelligence to remove the “noise” of false alarms, so that operators can focus on the true alarms that need their attention. 

What are true and false positives?

To avoid confusion, from now on we’ll refer to true alarms as genuine alarms, and to false alarms as nuisance alarms. 

When we’re training a machine learning model to learn the difference between genuine and nuisance alarms, we expose it to millions of images so that it can spot patterns of human and non-human activity. Ideally, we want the model to correctly identify what human activity looks like, so that it learns to ignore (or filter out) the non-human activity. 

As it processes an image, there are four possible outcomes that could take place: true positive, true negative, false positive, or false negative. Let’s look at how these work in the context of security camera alarms: 

As you can see, our machine learning model is making predictions about the image it’s analysing, based on what it has learned from the millions of images it has processed before. The two correct predictions are true positives and true negatives. We’re always trying to avoid false positives and false negatives.

False positives are the nuisance alarms that security operators are bombarded with day in, day out. We aim to reduce those so that they can focus on true positives - the genuine alarms that might need a rapid response. Currently, Calipsa’s False Alarm Filtering Platform reduces nuisance alarms by 92%. 

On the very rare occasions that Calipsa’s platform predicts a false negative, and misses a genuine alarm, we take full liability on behalf of our customers. Our model is currently 99.95% accurate at spotting genuine alarms, so we’re proud to say that incidents of false negatives are few and far between. 

So there you have it! A short and simple guide to true and false positives - and to true and false negatives too. 

Find out more about how our False Alarm Filtering Platform works, or see it in action with a free trial.