Very interesting.
We've switched from a mobile phone MAC address based location analytics people counter system(I believe founded by someone who worked on Google Analytics) to a stereo camera based system(bellwether).
We switched as our original provider's business model shifted(I believe in part due to the change in MAC addressing by telecom providers?).
We believe strongly in personal freedom and personal data security.
But we are scratching our head as far as what we can do to ethically and affordably better understand and serve our customers in an integrated(digital & meatspace) way.
Good Luck!
I'm not sure if our product would be able to meet your needs right away, but reach out to [email protected] if you'd like more information about what we have.
The depth map you provide doesn't contain the full output of the sensor, but how is that any different than using cameras but not passing the data up any higher? If you look at this video: https://www.youtube.com/watch?v=YOKMx7EDVys you can see the kind of image that Density has access to, though their depth map they show is only what's on the left side.
https://www.youtube.com/watch?v=WY_s6-WNZFU
https://www.youtube.com/watch?v=OBRFr1kkKkY
Hint: You need more than just 1 antenna ;)
Also the MAC address is obfuscated until a device fully connects to an access point.
Disclaimer: I may or may not have worked on such a system before.
Edit:
Fully open source MIT Licensed project that does it nicely by the look of it:
I was thinking more in terms of lidar tracking of people. It's been around for nearly a decade and works quite well.
Distance - I'd imagine that the distance from the top of a door frame to the top of the head of the average shopper is roughly equivalent to the distance from the Soli device to someone's hands.
Software - The Soli hardware streams radar contacts to software trained to separate and classify them, not too unlike the approach that Density is already using.
Benefits over a CV approach might come in power usage and lower complexity in the contact classification software.
Within spec, the algorithm does well tracking 8+ people at a time, and even when a line of people extends through a doorway we usually don't have to track more than that.
The hard thing about that approach might be handling conflicting signals and determining a "source of truth" in adverse situations. Generally speaking, there are variables in the real world that will trip up any given strategy (even if you were to station a human at each doorway with a hand counter). And while nothing can count perfectly 100% of the time, the difficulty is in finding an affordable strategy that will be really close or perfect most of the time.
Here's a neat demo of a multi-sensing device (not ours) which combines many signals to guess the activity taking place in a room: https://www.youtube.com/watch?v=aqbKrrru2co
We are a CCTV/deep learning system that uses existing infrastructure (think old school grainy security cameras), and we're are also able to capture additional information like age and gender of a person. Unlike other invasive/HD CCTV systems, we don't use facial recognition, and we also work over very wide areas, not just over restricted doorways.
MAC address tracking is a good simple way to get an approximate number of people: it's very easy to install, it requires only a WiFi antenna, and the data is easy to translate into a count. However, there are privacy and legal concerns which have prompted phone vendors to obfuscate this data. And there can easily be zero (or more than one) broadcasted MAC address per human, even when filtering by OUI.
Retailers have been using specialized "people count" technology like infrared break-beams, thermal cameras, and CCTV+CV systems for a long time. CCTV is the most accurate by far, but it's also commonly considered to be an invasive level of surveillance (and rightly so). It's also not particularly accurate in adverse situations. At Density, we looked at (and tried) many of these technologies - but ultimately found them lacking because they were either too invasive or too inaccurate. The device we've built uses a lower-resolution depth-only sensor because it can return extremely accurate results, without having the capability for facial recognition or other analysis techniques that are harmful to privacy. So far the technology is working very well - with an algorithm based on a deep-learning "human classifier" we're seeing accuracy above 99% in many of our deployments.
Here's a cool (and in-progress, excuse our dust) summary of some issues that the technology has to navigate: https://faq.density.io/algo/