How Density Works

Not-So-Standard Humans

Below is what our depth processing unit (DPU) sees in "standard entryways" all around the country. It turns out, people don't behave predictably. They do all sorts of odd things. 

They move in packs and linger.



They walk in different directions at the same time.



They peer through windows.



No two doors are the same. There are small spaces, and tall places, and rooms filled with light.



They hide beneath hinges.


And push between doors. 


People also like to bring objects with them. 



Backpacks and paper plates. Laptops and luggage.



Regular-sized trash cans.


Spaceship-sized trash cans.


And push carts. Lots of push carts.


We shouldn't have been surprised

Humans behave like humans. They like to hug and high five and they like to bring their kids. 


They wear hats and strange clothes.


If you're lucky, you'll get to see them dance.



Sometimes, they’re not even human. 



But no matter the size or shape they arrive in, we still have to count them; because all people count. 




One of the unique challenges in algorithm design for a problem like this is that when you account for one edge case, you often undermine another. The bigger your network, the more variation. The more variation, the more edge cases that compete for priority. 

Modern algorithm design: machine learning, support vector machine, neural nets, and deep tracking have given us new tools to address these issues of computer vision at scale. However, the gifs above are technically not flat images (like a camera). Instead, Density's data is three dimensional. As a result, there isn't a playbook. There isn't an open source solution. There's only what we build on our own.

To most people, counting people is simple. That's the problem with fundamental problems. They hide in plain sight. To us, counting people is as challenging as humans are unique.


Computer Vision

Density's machine learning algorithm, called "Metro," uses a unique classification system. Illustrated below is our method of precision tracking. The effect is achieved by separating a scene's objects into distinct categories.

In other words, Metro is simultaneously measuring the subject's head, hands, body, and any manipulated items while also keeping an eye on the door, surrounding walls, and other stationary objects that may be nearby. 


We hope this provides some clarity around the complexity of the problem we're out to solve and how our technology works.