The Measurement Minute by Gary Angel

 

One of the biggest challenges to doing people-measurement is the cost and hassle of the cabling. Deploying standalone single sensors has always been possible. But when data needs to be fused across multiple sensors to track customer journey, each LiDAR sensor must be connected via a home-run (a direct cable run) to a server in the IT closet that fuses all the point-cloud data together. That makes for a lot of cable and a lot of cost, especially in larger spaces or older buildings where wiring is often a real pain. Until now. Digital Mortar’s LiDAR distributed architecture solution lets you deploy multiple, fused, LiDAR sensors with no wiring.

 

The Measurement Minute by Gary Angel

 

One of our clients at Digital Mortar is a chain of gas/convenience stores measuring the full journey from pump to shop and back. What makes their use-case more compelling is that the stores are a significant upscale from your traditional gas convenience mart – with lots of more upscale food and coffee options. Those options likely increase the average ticket but also slow turnaround at the pump or in the lot – making measurement a great option for finding the optimal balance and tuning the indoor experience to maximize both dollars and efficiency. To measure the full journey, we need to measure from outside (the pump) to inside (the store) for each individual. Is that possible? You bet it is. And by combining multiple sensor types (LiDAR and Camera) not only is it possible, it’s efficient.

 

 

 

 

The Measurement Minute by Gary Angel

Although Digital Mortar isn’t really focused on loss prevention or security applications, there are two big reasons why it matters if you’re thinking about people-measurement. First, there are times when you can dual-purpose sensors to do both crowd-analytics and security (which greatly improves your hardware ROI). Second, when it comes to real-time people-measurement based alerting. the use-cases span operations, customer journey AND security. There are a number of basic perimeter monitoring and security applications that just fall out of good people-measurement.

 

 

The Measurement Minute by Gary Angel

 

Customer Journey alerting can completely transform the experience of a physical space. Instead of waiting for a customer to seek out help, it actively looks for opportunities to use the available help most effectively. Common use cases include identifying customers who absolutely require assistance (unmanned support, locked cases, etc.), and identifying shoppers in high-value areas where sending an Associate might have an impact. It’s even possible to look for evidence that a customer is struggling to find or figure out something (like a ticketing machine in a train station) and proactively send help. Perhaps most dramatically, you can flip the alerting paradigm on its head and direct Associates to the customer with the highest need or potential value, transforming alerting into a fully dynamic allocation system.

 

 

 

The Measurement Minute by Gary Angel

 

While performance may be the biggest challenge in building a people-measurement based alerting system, the most important problem for most businesses to solve is HOW to use people-measurement alerting. Use-cases tend to fall into one of these basic categories: operational management, customer journey, and security / loss-prevention. Of these, operational management is the most common and probably has the broadest set of potential uses.

 

Operational management use cases include but transcend line-management across multiple verticals. Crowd management, staff allocation, dynamically changing service strategies, providing public information for self-routing, and maintenance scheduling can all drive both ROI and better experience.

 

The Measurement Minute by Gary Angel

 

The visible part of an alerting system – the text alerts, webhooks and emails users get – is the easy part. The hard parts are building the realtime pipeline for store or location measurement, creating things like the predictive queue model, and keeping the underlying alerting engine performant so that Associates and Location Managers get the information they need quickly enough to use it. When it comes to people-measurement alerting, that last part is usually the crux of the problem. And it turns out to be a lot harder when it comes to alerts based on individual journeys (for things like shopper dwells, Associate interactions, and other over-time customer behaviors) than for things like queue lengths.

 

 

The Measurement Minute by Gary Angel

 

Productionizing a predictive queue measurement model really involves solving two problems. The first is integrating the predictive queue model into a real-time pipeline so that you can generate predictions about future line states based on current data. The second problem is deciding what to do about those predictions – how they get seen, distributed and acted on.

 

When it comes time to do the queue model integration, what kind of model you built, what you built it with, and how the rest of your pipeline is designed all matter. Productionizing can be fairly easy or fiendishly difficult!

 

The Measurement Minute with Gary Angel

 

Building a predictive queue management model is the penultimate and most interesting step in the process of building a queue management system. We used two different analytic techniques to tackle the problem: classic time-series and neural networks. The advantage of time series is that it is simple, easy to implement and productionize, and almost guaranteed to be performant. The advantage to neural networks is that we can advantage of some of the additional data we have about the store’s traffic and occupancy to improve the prediction.

 

The Measurement Minute by Gary Angel

 

Probably the most important step in building a model is making sure you’ve got the right target variable. Usually that’s not too hard. But looking at our line data, we quickly realized that predicting line depth wasn’t going to work as straightforwardly as we thought. The problem is that in some of the locations we’re working with, lines are actively managed. So we saw a roller-coaster pattern to line depths – where as the lines got long they would quickly come back down as registers were adjusted. In a managed line, you can’t use measured line-depth as your target variable. But there is a straightforward solution that works in both managed and unmanaged situations.

 

 

 

The Measurement Minute by Gary Angel

 

After the data wrangling comes the fun part of any analysis – building the actual model. That starts with model selection. For predictive queue management, we’re looking at time-series models. And because we have store traffic and occupancy data, it will be a multivariate model. In addition, it’s worth thinking about windowing and what model complexity before you jump in and start building something.