Crowd Analysis is critical for a myriad of applications such as retail, traffic, and event security. Current solutions require either large amounts of infrastructure or manual solutions such as count clickers at entrances and exits. We believe a more powerful solution with fewer resource needs was possible.

What it does

Analyzes video feeds to generate models of Crowd Density as well as the predictive direction of Crowd Flow over time.

How we built it

Data Analysis

Using node.js, we utilized Google Video Intelligence to create persistent trackers for objects in a given video feed. The resulting positional data was then analyzed to determine crowd density at a given time and then generate predictive direction vectors for the future motion of the objects in the scene. Image Recognition


Using Vue.js, we created an interactive visualization of our generated data. We applied data from multiple video feeds to create a heat map of crowd data and then place arrows in the direction of predicted crowd flow. This interactive layer was then overlayed on a map of on airport. Image Recognition

Challenges we ran into

Being able to accurately 'remember' an object between two frames of a video proved to be challenging. We experimented with several different detection models, but determined the Google Video Intelligence API to be the best fit. This was critical as it allowed us to look up the location history for a given object in a frame.

Accomplishments that we're proud of

We think Momentum stands out from existing solutions due to its predictive ability. To do this, we had to generate direction vectors for every object in a video feed. We then could analyze these vectors to draw broader, meaningful conclusions about the behavior of people and other moving objects.

What we learned

In building Momentum, we learned a lot of the various methods of video analysis and how to best apply that.

What's next for Momentum

We want to continue to make the predictive abilities of Momentum more intelligent by improving our direction prediction and adding more rich object annotations. We think that our work with Momentum is just a starting point. We are excited to see how people can come up with useful and powerful applications of this technology.

Share this project: