Counting people: It’s what Dor was created to do. Sounds easy, right? Surprisingly, it’s a difficult task to capture the flow of human beings in the real world.
Mike Lyons is a Senior Engineer who’s been with us since the beginning. Having built the brains of Dor from scratch, Mike is now responsible for making sure our machine learning algorithms keep getting more accurate and deliver ever higher value to retailers.
Here’s his perspective on the retail analytics tech landscape and his firsthand experience solving one of retail’s toughest engineering challenges: Capturing accurate customer counts without wasting power, time or money.
The main reason for Dor’s existence is that we’re geeks who got excited about solving this problem when we discovered that the other technology used for this purpose is either expensive, inaccurate or both. My team’s task was (and still is) to deliver the highest and most consistent accuracy in a variety of environments and doorways using the lowest amount of power in an affordable package.
It’s a tall order, and nobody has figured it out yet. We believe our approach gives us the best balance of accurate, actionable data, with deployment that's low-friction enough to actually make the hardware usable at scale.
The tech landscape
We set out to capture data in the physical world so that it’d be fast and affordable enough to be useful to the people who most need it. Our founder walked around San Francisco asking boutique retailers whether they knew how many people came into their stores. He heard an overwhelming theme: cheaper break-beams were inaccurate, and camera systems were $2000+ per doorway and therefore out of reach for a smaller business.
We went into business to solve this problem and were surprised to discover that larger retailers were having trouble rolling out camera systems to all of their locations too, citing installation difficulties and calibration costs as barriers to full organization-wide foot traffic visibility. It was a bigger problem than we initially thought, with an exciting opportunity to solve the people counting problem for retailers big and small.
Our team happened upon a thermal sensor after unsatisfactorily testing infrared and ultrasonic sensors for overhead people counting through doorways. This sensor detects light emitted by human heat. It’s counting the heat signature of the human head entering and exiting a frame. Our sensor uses a low amount of power that enables us to power our sensors with batteries that only need to be changed every two years.
We developed machine learning algorithms that count foot traffic anonymously and send those counts to our servers for processing into clean and accurate metrics. The machine learning running on each sensor in the field improves over time. When one sensor “learns” something, such as a new type of traffic event in an extreme weather condition, or a big crowd entering a store at once, we’re able to push that learning to all other devices in the field.
The accuracy our customers get from this setup continues to get better as our algorithm learns from their data. The sensors are also self-calibrating, so when installed correctly, you don’t have to tinker with them again. Because it uses batteries, installing the sensor is simple and doesn’t require hardwiring.
In addition, counts are anonymous, which can be a privacy benefit, especially in a healthcare environment where HIPAA compliance is legally required. Privacy is also becoming a bigger concern for retailers’ customers, so our sensor provides a time-saving shortcut past needing to get customer permission and gets you right to the big-picture trends without capturing any personally identifiable information.
This is how our sensor sees people moving through doorways, capturing whether it's an entrance or exit to count accurately, and protecting customer privacy through anonymity without comprising on accuracy.
Though thermal sensing has serious advantages, such and low power usage and protecting customer privacy, there is one key challenge we continue to address: extreme temperature environments. We teach our sensors to correct inaccuracies when they’re detected, and then push learnings to all other sensors in the field. Essentially, we have re-taught sensors to see people in extreme temperatures when anomalies occurred, and we can continue teaching them, meaning they’re always getting more accurate.
I chatted about this a bit in our latest webinar here:
To address this challenge, a project I’m currently working on is using an internal product for testing and confirming our accuracy. In the past, we’ve used manually entered data to regularly confirm our accuracy in the field, but as our business continues to grow, that system doesn’t scale.
We’ve been working with some of our awesome customers to confirm our sensor’s accuracy against a true bird’s-eye view of door traffic in different environments to get the best possible training data.
If we detect that our accuracy is lower than ideal in an environment, we can use the data we have collected to solve the accuracy problem in one of two ways. Most problems are solved by training our machine learning model on new data; the more examples of different problem cases that we can train the sensor with, the better all of our sensors become. Either we re-educate all sensors or change the framework itself to achieve higher accuracy. We don’t ever learn a lesson about our accuracy without teaching every sensor that lesson, too.
Where we’re going next
I’m working on an ongoing initiative for anomaly detection. We want to know ASAP when something outside the norm occurs in your data, mainly so that you can take swifter action. We want to be able to celebrate when you’ve successfully driven a flood of foot traffic to your locations! The other reason is being able to provide an even higher quality customer service experience when our sensor needs adjusting.
We take accuracy seriously here, which is why we chose a thermal sensor. It’s paid off by giving us a low-power, fast-deploying, cost-effective way to capture foot traffic metrics with room in our algorithms to reeducate and grow in value over time.
When it comes to in-store analytics solutions, cost and deployment speed are key for our customers, so as we look to the future, we’ll always work to give you the best option possible, meeting those concerns, without compromising on accuracy.
For more information on the types of people counting technologies out there, watch my 30-minute webinar on The Retail Analytics Landscape. If you're interested in doing this type of work with me, check our our job openings.