Choosing a People Measurement Sensor

A Decision Tree for Choosing the Right Technology for Flow Analytics

Opening up CCTV as a people measurement and flow analytics sensor is a big deal. Because there is so much existing CCTV infrastructure, it seems to promise a no/low capex, nearly universal solution for flow analytics and real-time operational monitoring. But that isn’t the reality of our world. In fact, the demands of video ML and real-time processing can make CCTV-based people measurement more expensive than buying new sensors and there remain significant issues around coverage, bandwidth and privacy. Having another tool in the toolbox is almost always a good thing, but it also puts even more pressure on the system designer to understand the tools most appropriate for each job.

With that in mind, late last year I built out a people-measurement decision flow for one of our transit clients to help make the right technology choice. As a transit operator, they’re (slightly) unusual in having a wide variety of use cases. They need to measure everything from very large stations to platforms to curbsides to individual trains. Their measurement needs and questions range from real-time crowd monitoring and queue management, to load-balancing, measuring trip-miles, understanding passenger journey, and dynamic maintenance. In short, they have almost every people-measurement use case there is and a lot of different spaces and locations (including rolling stock) to measure. So, they’re a really good test case for thinking about sensor technologies.

I played around with a lot of different ways to build out a decision-flow. In the end, the flow that worked the best started with a very basic use-case question: what are you trying to measure? If you’re primarily interested in counting people, you head down one major branch. If you need to track people (understand individual journey), you head down another. It turns out that your answer to this basic question sets the table for almost everything else. And while some sensor types show up on both major branches, their function and their ordering are very different depending on the basic type of measurement you need.

If you’re focused on counting, the technology decision tends to be fairly simple and is driven by space-size and existing coverage. In small,? confined spaces, measurement camera is hard to beat for price and accuracy. In huge open spaces, camera-based ML inferencing is about the only solution you have. And in mid-sized spaces, lidar sensors tend to shine.

?For individual journey tracking, the decision flow is more complex. In large spaces, the decision centers around the possibility and cost-effectiveness of continuous coverage. In smaller spaces, it mainly becomes a straightforward total cost of ownership decision between lidar and measurement camera. But full journey is inherently more complicated than counting. Journey tracking is never quite a binary, it’s more a spectrum. You may need to track some, most or nearly all journeys. Depending on the complexity and size of your space and where you live on that spectrum of some/most/all journeys, you may have to blend technologies together to get the tracking you want. That may include lidar and AI Camera rematch or even electronic tracking.

Here's the decision-tree in all its glory:

Where a decision-point requires a bit of explanation, I’ve numbered it. Here are the notes:

1. Crowd counting is usually a fairly low-value application and since large spaces will require multiple lidar sensors, that usually tips the economic balance in favor of camera with ML inferencing. It’s also true that the larger the space, the less value there is to the higher inferencing rate on lidar since the bigger the space the less variable the crowd size CAN be on a second-by-second basis.

2. Even in spaces that could be measured by a single lidar, if there are decent camera mount points or existing camera infrastructure with good views, using camera is usually less costly and should provide a “good-enough” solution in most cases.

3. Measurement cameras are ideal for choke-point measurement if you have a good top-down mount point and stable lighting they are nearly always the best solution.

4. ML inferencing is expensive and processor intensive – making it particularly challenged in high-volume cases. But if you have existing camera coverage, ML inferencing is an option especially for one-time or short-term deployments (like Live events).

5. Even very large spaces MAY allow for continuous coverage. If the space is unbroken and has useful mount points for sensors in every area, then it may support good quality full-journey tracking without any form of re-match.

6. If continuous coverage isn’t possible and there are NOT defined endpoints, then lidar sensors for the coverage areas are probably the best possible solution. Note that this will not provide true full journey coverage across the space and may limit the use-cases that can be reliably delivered.

7. If a space doesn’t have continuous coverage but has well-defined journey endpoints or chokepoints (think of an airline terminal with gates or exits being endpoints), then camera re-match can provide full-journey tracking albeit with black-out areas along the way. Blending lidar & camera re-match in this situation will usually create something close to true full-journey tracking.

8. For small spaces where camera is practical, the decision usually comes to Total Cost of Ownership – price. Measurement quality between camera and lidar is roughly equivalent so whichever solution provides a lower TCO is probably the best choice.

?

The decision-tree doesn’t capture every facet of sensor choice. You’ll notice, for example, that privacy isn’t addressed at all. If your organizational guidelines forbid using camera for tracking, not all the branches on the tree are available to you. Location specific limitations on wiring or bandwidth can also shape sensor choice.

For the most part, though, the decision-tree will get you thinking about people-measurement sensor technologies in the right way – helping you match your choice of technology to the use-cases you have and the types of location you need to measure.