In retail, footfall-related metrics have become cornerstone data points used to track various aspects of store performance. Everything from RoI of marketing spend, to staffing requirements and, of course, in-store conversion rates.
That being said, after having hundreds of customer conversations, we’ve noticed a trend: based on systems they've used in the past, many retailers have trouble trusting the data, specifically due to problems with accuracy. Ultimately, these fears almost always end up with a lot of value being left on the table.
As this seems to come up frequently, we thought it would be useful to write an article detailing the accuracy expectations across a range of different people counting technologies.
It’s worth bearing in mind that different technologies are useful for different applications. However, shining a light on where inaccuracies lie can help you interpret your people count data, or assist in choosing which technology most suits your current application.
In the following section, we will discuss the commonly available technologies, describing how they work, and what will cause them to fail.
This technology ‘sniffs’ out mobile phones by sending wireless signals within their detection range, and infers a count number based on this. There was some hope for this approach in the mid- 2010s, as smartphones became ubiquitous, and ‘big data’ was the buzzword du jour.
However, the tide has receded for this approach, as fundamental barriers have not been overcome, including sample variances (e.g. some people have two phones), MAC address randomization which can make one phone appear as multiple phones over time - and proximity detection problems - where the technology cannot distinguish between someone just inside or outside a doorway.
As such, this method is useful for a specific set of circumstances, such as in gaining a broad understanding of visitor paths in a very large area - but is not appropriate to collect data on a specific number of store visitors, or for use in conversion rate calculations.
This type of counter uses a beam of infrared light pointed at a receiver or reflector, counting crossings whenever the beam is broken. Their main advantage is that they tend to be very cheap. They can also be battery-powered and usually allow for quick and easy set-up.
Unfortunately though, in almost all cases, they have very low accuracy. These inaccuracies are in general caused by fundamental flaws in the crossing detection method. As count is measured whenever the beam is broken, people passing through the beam side by side or in a group will often be undercounted.
On the other hand, if someone walks in and out of the doorway multiple times in a short period, or if something other than a person breaks the beam, again these counters will lose accuracy. In addition to this, they're difficult to audit, monodirectional, and need to be placed in a location where you'll be able to also mount a reflector/receiver.
Optical counters come in all shapes and sizes - from monocular to stereo, depth-sensing, or thermal. They are the most accurate category of sensors by far and can be used for indoor or outdoor counting purposes.
Thermal counters use an infrared camera to capture a heatmap of their environment. One of their main advantages is that, as they're not capturing visible light, they are naturally GDPR compliant. They also work well at any light level however, during regular retail business hours, this is rarely utilized. A number of factors can affect their accuracy, for example, heat-reflective surfaces can mask people from the view of the camera, and the video they produce is more difficult to audit.
Stereo counters are popular and work by detecting depth much like humans can, and do a good job in distinguishing between a human and, say, a bag. That being said, they do tend to be expensive and complex to install, but they generally provide a good accurate count if installed correctly.
Monocular cameras in the past have suffered from many accuracy problems- they are built to count any object that moves beneath them - which can set off a count even if it’s not a person. For example, a pram, or large bag, could be counted. Furthermore, the early simplistic tracking algorithms resulted in miscounts or multiple counts of people, as unusual movements or dwelling triggered the count.
These days though, the issues with monocular cameras have been solved by the recent advances in machine learning and computer vision algorithms. The latest algorithms coupled with good hardware can overcome almost any environmental or behavioral issues. At HoxtonAi, we retrain our algorithms almost every week, to ensure the highest data quality across our client base, which means we don’t need expensive dual-camera setups to achieve high accuracy rates.