How We Built an AI-Powered Footfall Counter

How We Built an AI-Powered Footfall Counter

In an intensely competitive retail environment where change is not just consistent but exponential and disruptive, staying responsive to evolving customer buying journeys is critical for us and our partner customers. Products can no longer be about simply creating a temporary fix to a market need but providing transformative, long-lasting solutions.

Although at Capillary we were capturing a lot of omnichannel consumer data for our customers through our retail CRM solutions, the data captured in offline stores was purely post-transactional in nature. Brands had no way of identifying potential customers among store visitors or optimizing their in-store experience proactively.

This was in sharp contrast to online retailers, who had easy access to pre-transactional customer data which included everything from how many people visited their online store to their browsing behavior, time spent on-site to pages most visited, even cart analysis, abandonment and final purchase.

These insights enable online retailers to influence and tweak customers’ shopping experiences in real-time, positively impacting sales.

Recognizing this, we sought to build on our existing CRM platform by developing a solution that could give stores similar strategic capabilities in an offline model. By proactively personalizing buyer journeys based on advanced customer identification, customer visual profiling, staff optimization and deep store insights, stores and brands would be empowered to positively impact in-store conversion rates.

This was the vision our team was entrusted to work upon.

To this end, our team crafted the larger Instore Vision’ which projected the creation of ‘Smart Stores’ or digitally enabled offline stores where in-store shopping experiences could be customized by infusing online elements such as ‘Virtual Store Staff Assistants’. The entire Instore Vision was rooted in the basics of store performance i.e. Store Traffic Analysis. This entailed focusing and analyzing the very start of a customer’s in-store buying journey, to answer a question as old as the bricks and mortar that are used to build the stores themselves.

“How many people have visited my store?”

Customer footfalls, or store traffic, is by far one of the oldest and most important KPIs of any store.

The number of customers that visit a store is the first indicator of a store’s performance and potential sales. Conversion Rate, a key store performance indicator, derives wholly from footfall traffic.

What we realized, however, was that this data was woefully unavailable and inaccessible in any reliable, structured and, most importantly, accurate manner.

In most cases, data on a store’s foot traffic lacked transparency and was purely anecdotal, based on assumptions and estimates. This outlined a clear need for a footfall counter that could accurately capture the number of visitors at a store and process this data in real-time in a reliable, structured and actionable way.

Woman under stress

Offline retailers often stress out trying to count footfall

Capillary vs. the World

Footfall counters are not a novel idea, yet stores and brands still grapple to assess basic footfall data.

We first had to research, in no small detail, all the different types and models of footfall counters available to retailers in the market. These finding helped us define the final product and our targets of achievement over the next few months.

We learnt that in order to be an absolute value-add to retailers, our footfall counter would have to be –

  • Cost-effective and scalable
    Our focus towards the Asian market meant we would have to make our footfall counter more affordable than existing options to deliver a positive ROI and not hurt bottom-lines.
  • Over 95% accurate with no cases of over-counting or false-positives.
    As against the 70% accuracy standard (actual) of many existing solutions that often resulted in cases of over-counting (counting the same visitor more than once) and false-positives (incorrectly identifying store staff or non-human objects as visitors), rendering them ineffective and unreliable in gauging store performance.
  • High on availability with 100% uptime and seamless connectivity
    To combat the issues of downtime due to outages on existing power-run, Ethernet-connected options which sometimes led to incomplete data being captured.

We knew that achieving this would by no means be an easy feat, as we were going down a path riddled with challenges that many have tried to take on, but in true Capillary spirit, we were excited to make it happen.

With all systems a-go, our footfall counter, VisitorMetrix was scheduled for release in January 2017.

Breaking the Mold

Existing solutions which ran on multi-layer background extraction and blob detection technology delivered highly ambiguous results, being incapable of differentiating between human and non-human objects and were easily affected by changes in lighting conditions.

It was clear from the onset that existing technology would not allow for a product to be built on the lines that Capillary had hoped to achieve. We concurred that artificial intelligence would be the key differentiator that would set our solution apart and provide both the company, and its customers, a strategic competitive advantage.

With AI adoption in the retail industry still in its nascent stages, developing in-house, AI-based programs on which we were to center our first hardware solution was a bold move. However, given the ready and visionary talent at hand, we ventured on with no trepidation.

Behind the Scenes

Subrat Panda, Principal Architect and key driver of Capillary’s company-wide AI initiatives was pulled into the project right from the beginning. With a PhD in Computer Sciences, Subrat is representative of the human expertise within Capillary that drove the adoption of artificial intelligence in the development of VisitorMetrix.

Unlike other providers who banked on third-party software solutions, Capillary had a clear objective to not just be a reseller but a proprietor of the technology we developed and utilized. Subrat and the other architects in our team went to work ensuring this by enhancing existing blob-detection technologies with advanced machine-learning algorithms. The team worked tirelessly at refining the algorithms VisitorMetrix would use by extensively training them on over 12 million images. This was to enable blog-detection programs to recognize visitors to a store in a more ‘intellectual’ way.

In order to do this, we didn’t follow the norm of depending on third-party, open-source libraries to obtain images but went about creating our own datasets based on actual store visitors and conditions for highly accurate results.

Doubts and turmoil

With only one month to go for the deadline, a status check revealed that while VisitorMetrix was checking the right boxes in terms of cost-effectiveness and connectivity, but it was still not hitting the mark on its key success target – accuracy.

At accuracy levels between 60% – 70%, a large number of false positive results were recorded, which meant the team was not confident about obtaining the desired IP.

The release date was consequently revised to April 2017, giving us only a 3-month leeway to identify issues with and rework the algorithms that would make VisitorMetrix effective.

The delay in release created a moment of uncertainty for us, and made me contemplate the improbability of meeting the revised deadline and making up for the loss of potential revenue an early release would have entailed.

A shopping cart

Pictured: A human according to many blob detection algorithms

Ghosts of products past

VisitorMetrix was, incidentally, not our first foray into developing a hardware solution. The team’s first attempt at a hardware product was in 2015 with a Wi-Fi beacon which was to be used for geo-location, in-store customer identification and in-store engagement. Despite having invested heavily in hardware and achieving the results we hoped for, ultimately the technology we used had become redundant by the time the product was ready. iOS devices were now transmitting scrambled MAC IDs, hence we were unable to identify store visitors. The entire Wi-Fi-beacon market had tanked by then, making our product unviable even before its launch.

This incident, however, was just a chink in the Capillary armor. Far from being deterred, the team welcomed the challenge to roll the dice once again with VisitorMetrix. This time we knew we shouldn’t build a product that relied quite heavily on other third party systems but build as much capabilities as possible in-house to have the most control over product performance.

The AI training montage

Albeit having been demoralized by the delayed launch date, we ploughed on believing an accuracy target of 95% was achievable by fine-tuning the way images were captured and processed. We revisited our previously captured footage through the auxiliary camera in the device and after careful analysis it was revealed that in-store conditions such as lighting, shadows, reflections or objects in the external store periphery such as trees or cars for example, still affected the way images were captured, resulting in skewed accuracy results.

Having identified the glitch, we decided we had to ditch blog-detection completely. We were not going to achieve the accuracy we were hoping for with this. The team now rallied together to rework the algorithms, now relying on computer vision instead of blog detection or background extraction. With machine learning algorithms we were certain we’ll enable the system to accurately differentiate between human and non-human objects such as carts, shadows, etc.

We set up devices at 40+ stores to capture real-life visitor images. This included auxiliary devices to capture walk-ins and walk-outs that together yielded over 12,000 hours of video footage. Not only was this footage impressively extensive, but varied.

VisitorMetrix is trained to recognise just human images, for now…

Over 3 million images of this database were used to train the algorithms to identify humans from the top-view and 9 million impressions of non-human objects such as shopping carts, bottles, helmets etc. (that normally distorted results) were used to train the algorithms to differentiate between human and non-human objects.

Further rounds of intense image training followed until the algorithms were rendered capable of identifying visitors to the store accurately irrespective of store conditions. Now that the IP was almost within reach, it was imperative we didn’t leave anything to chance.

We were determined to make the April release an absolute success. The team, which is essentially an in-house unit, took dedication and collective effort to a whole new level when they decided to go into stores themselves to test and ensure product viability for customers under real-time circumstances.

Accuracy Realised

This exercise in perseverance put us on the right track and at the close of the three-month period, VisitorMetrix had achieved an astounding accuracy rate of over 95% having eliminated all cases of over-counting and false positive results.

This was the technological breakthrough the team had been working for and after 9 months of sheer dedication, VisitorMetrix was on schedule for an April release as Capillary’s first fully AI-powered hardware product.

The team didn’t rest there though. After having successfully developed and launched VisitorMetrix, the team immediately switched gears to ensuring the footfall counter was not just technologically but also market viable as well.

The 6 months following the release saw the team focus all their efforts on stabilizing the product to ensure optimal functioning under real-time conditions and fine tuning the device to account for full-functionality on all counts.

VisitorMetrix team working over the weekend to ship the first batch of devices. From left, Saurav Behera ( Senior Firmware Engineer), Sumandeep Banerjee (Principal Engineer) along with the first interns of the InstoreAI team, Harsheel and Ruchita

 

VisitorMetrix – The Launch

VisitorMetrix was officially launched to the market on October 8th, 2017.

At the time of its launch, VisitorMetrix was exactly what we had envisioned it to be viz. an AI-powered footfall counter that was–

Cost-effective – At a unit cost way under 250 USD, VisitorMetrix incurs very low set-up costs, making it both an affordable and scalable option for stores across all models.

  • Accurate – VisitorMetrix provides over 95% accurate results with zero cases of over-counting or false-positives which was important for us.
  • High on uptime – VisitorMetrix offers 100% availability or uptime via WiFi or LAN networks for uninterrupted connectivity.
  • Low on maintenance – VisitorMetrix requires zero maintenance and is easily configurable with remote troubleshooting

Looking Ahead

Since its launch, VisitorMetrix has found a home in over 300 stores, across 40 cities and its growing demand stands testimony to its commercial viability.

VF Brands and Promod were among first major retailers to deploy the footfall counter to their stores. In our webinar, “How AI can bring power back to stores”, Pankaj Agarwal, Retail Director of VF Brands  spoke to the veracity of how VisitorMetrix drove a notable increase in store sales, based on customer traffic patterns.

Even as VisitorMetrix continues to gain market traction, our teams have already moved on to realizing the next phase of the Instore Vision offering, buoyed by the success of creating Capillary’s first AI-powered hardware solution; a win for both Capillary and our customers.