Clever crowd counting: fast, accurate and helpful

Look at the crowd above. How many people would you say were in the photo? How long do you think it might take you to count them? And if it was just a small part of a bigger crowd, what is the calculation to estimate how many people were there altogether? It’s a tough call, isn’t it? So, you now have a little insight into the way crowd counting used to happen – arduously, undertaken by humans and, as a result, somewhat lacking in the accuracy department. Today, however, things are pretty different.

Crowd counting is important. We all want to be able to live our daily lives in a way that is safe and convenient and knowing how many people are in a given space at any time can sometimes be essential to this. People in charge of spaces and places need to manage the movement of large numbers of people, as well as plan the resources needed to look after them. So, where there are huge crowds of people – concerts, sports events and festivals, for example – there is crowd counting. But it’s also critically important for public spaces, such as airports, train stations and shopping centres. As you might imagine, these are places where manual counting just isn’t ideal.

Back in 2016, Canon released a piece of software called People Counter, which uses video content analysis technology to count the number of people present in images captured by network cameras. Later, in 2019, this was followed by an updated version (called Crowd People Counter) which was able to count thousands of people in seconds, thanks to developments in Artificial Intelligence. A proof of concept at an international rugby match in 2018 showed that it could count around 6000 people in just a few seconds with a margin of error below 5%, when compared to manual counting.

You’d think that simply ‘counting’ would be fairly straightforward for a piece of software to achieve, but it has its complexities. For example, one of the biggest challenges lies when people ‘overlap’, standing in front of each other or too close together. For a piece of software to differentiate one individual from another can become tricky. To solve this, Crowd People Counter uses AI to detect and count only the number of heads, not faces or bodies, in a crowd, which is more precise. Of course, this meant that Crowd People Counter’s algorithm needed to know what a human head in looks like in all possible crowd situations and at every angle that a network camera might view them. To do this, it was ‘fed’ a huge volume of example images, with the heads marked, then the developers watched to see if the algorithm’s accuracy in spotting them increased. Sounds simple, doesn’t it? Not when you discover that this process involved giving the algorithm several hundred thousand computer-generated 3D crowd scenes – a number that would be near impossible to come by otherwise.

The process was, of course, a learning curve over time that required both hardware and software to advance and imaging technology alone has developed a great deal in just seven years. However, huge progress in Artificial Intelligence meant that crowd counting technology went down the path of deep learning early on. A subset of Artificial Intelligence, deep learning mimics the human brain to solve complex problems by recognising patterns in data – much in the way we humans do when we see the world around us. So, to boost the precision of Crowd People Counter, the team developed a lightweight deep learning model, not requiring the seriously heavy processing power of some models and making it far more efficient to run.

One of the key things that makes Crowd People Counter different, however, is the way that the development team behind the software were able to tackle challenges. They had camera development teams easily available to speak to, which helped them when they had issues of image noise under low light, for example. As a result of this kind of happy collaboration, the software now achieves a high degree of precision. Equally, working across Canon family companies, with Axis and Milestone, gave everyone a way to realise some valuable mutual capabilities, such as real-time trend analysis and low-distortion high-megapixel image resolution. As a result, Crowd People Counter finds itself applied across a variety of fields – for plenty of different purposes.

Which leads us neatly back to the ‘why?’ It’s obvious that crowd counting can keep us safe. It can prevent busy situations turning into deadly ones, simply by alerting security personnel to the potential for dangerous overcrowding. It can make sure that access to spaces is measured or help us to understand the maximum amount of people who can visit a place safely. There are also some quieter, everyday ways that crowd counting can make things a little bit better. It can show us helpful trends – how many people might be expected to visit a restaurant, for example. This can help businesses plan how much food needs to be prepared on any given day, which can considerably reduce waste. It might also be used by town planners to understand how spaces and places are used, and then apply this knowledge to encourage more environmentally friendly transportation options. In the days before such technology, this would be a really time consuming and potentially complex task, but today there’s no need to use just eyes and a calculator. Unless, of course, you want to.

Learn more about Canon Crowd People Counter.