This TfL AI experiment reveals how Tube station capacity could be increased – without building anything new

Incremental gains ftw

Bloody tourists who don’t know how to use a ticket barrier.

The most annoying thing about the HS2 debate is that some people still don’t know what the new railway line is actually for.

This is partly the fault of branding. Even though it’s called High Speed 2, the core purpose of the line is (or was supposed to be) to increase capacity. If we can relieve the congested West Coast Mainline that connects London with Birmingham and Manchester, then more people and more freight can travel via rail.

Presumably though, at some point it was decided that “capacity” isn’t a very sexy selling point, hence the messaging emphasis on the minutes the line would save when travelling on it.

It’s a shame really, as capacity is probably more important when running a transport system. Higher capacity means more customers served and a more reliable service.

For example, the Elizabeth Line in London will get you from Paddington to Liverpool Street faster than the alternatives, but it would still be useless if it only ran once an hour. The real value is that it has 24 trains-per-hour, each capable of carrying 1500 people. And that it means you don’t need to travel on the horrible Central Line.

So from a transport network design perspective, capacity and the throughput of passengers matters, and it is definitely something to worry about, because capacity bottlenecks can have network-wide downstream consequences.

Take the below map for instance. On paper, and in the fevered imagination of every spoddy man inside the M25, it seems obvious that Transport for London (TfL) should connect up the mainline Clapham Junction Station with the Tube’s Northern Line. As now the Northern terminates at the new Battersea Power Station redevelopment, it is painfully close to joining up¹ with the UK’s busiest interchange station².

However, the reason it hasn’t been done is because connecting directly with Clapham would overload the Tube. It would direct thousands more commuters on to the Northern Line, and there simply isn’t the capacity to cope with the extra influx.

And what makes this connection an even more unlikely possibility is that untangling this capacity crunch would itself be a nightmare.

In theory, there is a way to do it. It’s could be possible to run more trains on the Northern Line by splitting it in half. Londoners reading will be aware it’s a strange line as it contains two almost completely separate branches – with one half running into the City of London, and the other half through the West End, with the lines converging at Camden Town – before splitting again to head on two different branches into North London.

So in principle, the almost-completely-separate two branches could be severed completely into two separate Tube lines, meaning that trains can run on both more frequently – increasing the line’s capacity.

But this isn’t easy to do either, because of another capacity issue: If the line was fully split, Camden Town station – where the two branches meet – would be similarly overloaded with passengers interchanging from one branch to the other. So this means that before Clapham Junction and Battersea Power Station in South London can be connected, you first need to bulldoze and rebuild a major station in North London.

In other words, its capacity bottlenecks all the way down. And the elephant in the room is obviously money: The reason these are bottlenecks is because rebuilding and enlarging stations hugely expensive.

For example, when Bank was enlarged a few years ago to create more space on the Northern Line platforms for passengers to wait, it cost around £700m. And upgrading Camden is estimated to cost a similar amount.

So the obvious question to ask is: Is there any way to upgrade station capacity without an expensive rebuild? If we can squeeze more people through the existing stations, that’s better for everyone – from passengers wanting a comfortable journey to TfL’s Chief Finance Officer not wanting to have an aneurism.

Excitingly, I’m pleased to say there is – at least, to an extent.

Using the Freedom of Information Act, I’ve obtained some internal TfL documents that reveal how Tube station throughput – literally the number of people passing through one common station bottleneck – could be increased, meaning that conceivably more people could be squeezed in across the Tube network.

To be clear, this likely isn’t going to be a world-changing capacity improvement. This isn’t like how adding the Elizabeth Line singlehandedly added 10% to Tube’s capacity. It will not make it possible to connect Clapham and Battersea. Nothing will beat bricks and mortar for that. We still need to Keep Building Stuff.

But this is a very cool incremental gain that, if TfL were to roll it out, could squeeze that little bit of extra juice out of the existing Tube network.

So let’s dig in and find out what it is.

If you enjoy nerdy stuff like this, then subscribe (for free!) to get my takes on politics, policy, media, tech, and – of course – transport, direct to your inbox:

The intelligent gateline

Our story begins at Blackhorse Road Tube Station, at the northern end of the Victoria Line in Walthamstow. It’s primarily a commuter station serving a residential area – and every day the passenger flow is what you’d expect. In the morning, most people are heading into the station, en route to work. And in the evening, it’s the reverse.

On their way in passengers will, of course, have to pass through one of the seven ticket barriers - or the “gateline”, to use TfL jargon.

Typically gatelines are a major chokepoint in a station, as everyone has to pass through a gate one at a time, tapping their Oyster card or phone on the way.

Remember these seven gates as they will become important in a minute.

To get as many people through the gates as possible, they are designed to be configurable. Station staff can choose which are open and in which direction. This means that, for example, you can have more entry barriers in the morning, and more exit barriers in the evening to match demand.

However, changing the direction of barriers is not something that typically happens very often. Staff might switch them over at a set time of day. Or if they notice a build up of people queuing, perhaps they’ll switch a gate over manually. But as things stand, judging by the documents I’ve obtained, it is not a particularly dynamic process.

And we’ve all been there, silently swearing at the tourist ahead of us, as they fumble with their phones and stop dead right at the barrier.

So you can probably guess where this is going: What if the gateline was more responsive to real time conditions in the station? What if it could automatically swap the direction of gates based on where the crowds are coming from? If the gates could flip at the right times, that means increased station throughput – and thus more capacity for passengers inside the station and across the network. How much more efficient could that make the Tube?

This was what TfL set out to find out, enlisting transport tech company Cubic and the University of Portsmouth to help.

If you haven’t subscribed yet and have read this far into an article about the efficiency of ticket gates then what on earth are you doing? You’re clearly part of my core demographic.

How it works

To make it all work, TfL first needed to figure out which direction passengers were coming from. To do this, it used cameras³ made by a company called Xovis that are designed to track and count people, figure out the direction they are travelling, and how long it takes to queue⁴.

Here’s what one of the cameras saw:

And here’s a representation of how the cameras turn images into data: You can see where all of the individuals in the ticket hall are (the blobs), and the paths they have taken to get there (the lines trailing). The yellow boxes are where queues may build up.

Most boring Metal Gear Solid VR mission ever.

Next it was then simply a case of… letting the AI take control of the gates. But here, TfL took an ultra-cautious approach.

For example, as far as I can tell from the docs, the two extra-wide gates at either side were kept on manual mode for the duration. It was only one gate – number 42 on the diagram above – that was ever operated in fully autonomous mode, switching automatically based on what the sensors were saying.

The other four (40, 41, 43 and 44) were operated in “supervisor mode”. This was where the Cubic system would decide when directional changes were needed, but instead of just flipping the gates over, it would first send a request to a tablet held by a member of staff – who could then accept or decline the change.

Action shot of a member of staff using a tablet to change direction.

So in essence, the actual hardware deployment was relatively simple. Though behind the scenes there was a lot of hidden complexity.

For example, TfL and Cubic enlisted the help of the University of Portsmouth to figure out the optimal mathematical model for triggering changes to the gateline. Here’s some complex-looking probability calculations⁵ that gave me anxiety flashbacks to A-level maths:

The model was also apparently designed to be agnostic on the type of sensors used too – it’s conceivable the same monitoring could be achieved by, for example, sniffing out bluetooth signals from passengers’ phones instead of using AI cameras, or by plugging into existing CCTV (a bit like the later TfL Smart Station trial). But these other techniques were not tested⁶.

And one other interesting technical tidbit is that the sensors were also set up to have a safety break: The cameras also kept watch of the top of the escalators (that’s the big yellow box on the diagram above), and it would alert the staff to close all of the gates if it was getting particularly busy, to help manage the crowds.

This is clever as it shows two things: First, the potential safety benefits from a system like this, but also why improving throughput is important: Because apparently in just the two weeks the hardware was tested at Blackhorse Road, apparently there were eleven “escalations” with the safety alert, suggesting stations can become dangerously busy with relative ease.

Looking at the numbers

Given how it was designed, with most of the gates merely in supervised mode, it was a relatively timid trial. If I were in charge, I’d have recklessly set all of the gates to automatic – but I guess TfL is more responsible than I am.

But even given these limitations, the results are striking. To work out the real impact this tech could have, TfL fed the real world data into a lab-based simulation model, which assumed all of the gates were controlled automatically. And it reveals that the impact could be significant.

For example, the study found that this AI could reduce queuing times by something like 75.6% or even 89.63%, depending on which mathematical queuing model was used (hence the complex maths above).

And for residential stations like Blackhorse Road, the study concluded that it could increase passenger throughput at the gateline by around 30%⁷. That seems pretty dramatic to me! Think of all of the extra passenger capacity that could be squeezed out of the existing Tube system – especially during peak times.

Barriers to adoption

Given the apparent advantages of using AI like this, you might think that the next obvious step would be to role out the technology across the Tube network.

For a relatively small, fixed-cost investment in each station, TfL could scale this efficiency gain across the five million passengers journeys a day on the Tube network, which adds up to a network with more capacity, fewer bottlenecks, and fewer annoyances during rush hour.

And what might surprise you is that TfL has had plenty of time to do this: The trial took place between 2017 and 2019. That’s ages ago now – for context, back when the project began, Peter Capaldi was still Doctor Who, Elizabeth Holmes and Sam Bankman-Fried were still inspiring business leaders, and I still had a full head of hair.

However, since testing at Blackhorse Road, nothing has happened. According TfL’s response to my request, “since the completion of the trial in 2019 we have not progressed the project further.”

Unfortunately I’ve no further information on why TfL has left this compelling study just sitting there, but if I had to speculate, I’d bet that it’s a combination of the extreme financial pressure that TfL is currently under, and the fact that just a few months after the trial concluded, whoever was championing it at TfL presumably had bigger, pandemic-related things to worry about⁸.

In any case, this seems like a missed opportunity not to do something further with it⁹.

In the longer term it seems likely to me that even if passenger numbers haven’t quite recovered from the pandemic yet, there is a structural inevitability that at some point in the future, capacity could become an even bigger issue for London’s transport network.

Why? Because pick any Tube station at random and chances are you’ll see some pretty dense residential development in the immediate vicinity¹⁰. For all of the NIMBY problems, London is still steadily building new homes near Tube stations, and this means more people to squeeze into the existing Tube system.

So it’s inevitable that capacity will continue to be an issue. And while we continue to dream of the day we can one day take the Northern Line to Clapham Junction, this at least seems like a pretty neat way of making more of what we’ve already got.

Phew! If you’ve read this far then you should sign-up (for free!) to get my nerdy takes direct to your inbox:

Also, don’t forget to follow me on Twitter:

And because you clearly enjoyed reading this, you will definitely enjoy my TfL “smart station” story from back in February. Go read it!

1. This is completely irrelevant and tangential but another really satisfying bit of Tube-map-joining-up would be to extend the DLR west from Bank. Imagine if it were to go all the way to the old Jubilee Line platforms at Charing Cross (where the Jubilee terminated before it was extended south of the river in 1999). You could even have extra stops at the old abandoned Strand/Aldwych station, and City Thameslink, finally putting that latter station on the Tube map. Sadly, however, even if money were no object it wouldn’t make much sense, as it would just mean that by the time trains got to Bank they were already be overloaded – defeating the point of a railway designed to take people from the City to the Docklands.

2. I understand from talking to men in anoraks that the developers of Battersea Power Station have even constructed the new station in such a way that it would make one day building the extension more viable.

3. In my experience covering these things, the companies that make these sorts of devices are very, er, sensitive, about their tech being referred to as a “camera” for obvious “inviting people to complain about privacy” reasons. So I’m sure the company and others like them would rather I call it a “sensor” – and to be fair, there are some differences: Though there is an optical sensor (like a camera), images are processed on device, and though I’m not sure exactly which model of sensor used in this trial, it appears that Xovis devices are configurable to not share images, and only spit out data.

4. I’m not sure if the capability existed in the cameras used in the TfL trial – this was a few years ago – but the cameras the company currently sells are apparently capable of tracking which way customers are looking. In other words, as you look around the supermarket shelves, this technology can figure out what products are catching your eye. Crazy.

5. Well done inevitable smug man in the comments who intuitively understands this, have a pat on the head.

6. When the trial was first designed, there was a plan to test some of these other technologies at Marylebone station – though this was apparently abandoned because of the complexity of installing the necessary kit in a listed building.

7. Apply this model to other types of station, such as big mainline terminals or stations in business areas may be different – or was at least out of scope of this model.

8. And I’d bet that it hasn’t been picked up again not just because of the aforementioned financial pressure, but also because it isn’t clear yet how long term Tube usage/commuting patterns will have changed. Tube usage is still only at around 75-80% of what it was pre-pandemic, so perhaps throughput simply isn’t as pressing an issue.

9. It is possible to imagine how there could be even more efficiencies lurking in the data too if they continued to build on it. For example, imagine if the gateline was plugged into real time transit data. When a bus parks up outside or when a train pulls into a platform, the gateline could quietly configure a few extra gates to prepare for the brief glut of foot-traffic that will be quickly approaching.

10. For example, take a look at Tottenham Hale and Bromley-by-Bow, which are both residential stations that are seeming nearby former industrial land turned into flats. Over the last few years Finsbury Park has transformed too, with a major residential development almost directly on top of the station. And Elephant and Castle is the same.