Sofa Mobility Report: Building mobility analytics similar to Apple and Google

The archive of snapshots that was available at the moment when this bright idea came across our minds was, unfortunately, a little bit out of date to capture the big drop in mobility caused by increased strictness of social distancing rules in late March. So, we analyzed the available timeframe: from the 16th of April till the 8th of July, almost three months.

The technology behind this analysis was OpenCV, YOLOv3, and Tensorflow — the open-source libraries and frameworks available from the shelf. In terms of the people detection, we fully relied on YOLO which is the state-of-the-art (best in the block) machine learning model for object detection.

Our calculations for all the fifty cameras brought us the number of people that varied approximately from 5,000 to 29,000 per day. The dotted line is a trendline — a linear approximation of our jumping daily numbers. As you may notice, there are several dates where the data is missed, those are some technical issues on our side or just the days missed in the archive.

CCTV: people per day

How reasonable are the numbers we got? Well, more or less when we recall that we process only once-in-5-mins snapshots. This is 12 images per hour. In reality, people appear more often in front of the cameras, so ideally we would take a frame roughly every 20–30 seconds, which would make us 120–180 frames per hour, 10–15 times more than we have. Multiplying it by our numbers, we would get 50,000–75,000 people on an empty day and 290,000–435,000 on a crowded one during the times of pandemic. Seems legit for a 5 km circle in the center of London considering limited cameras view.

An example of the camera stream in Central London

Apple allows downloading the raw dataset so that you can play with the numbers compared to the baseline which is 100% by the y-axis (and they also lack a couple of days). 33% value on the chart means that Londoners walk less for 67% en masse.

Apple: mobility drop below the baseline

While the Apple report describes the whole of London, the camera analytics observes only a part of central London that consists mostly of offices and public spaces having fewer residential buildings. So, our assumption was that while people living in bedroom communities continued walking to the grocery stores, parks, and other venues, the traffic in downtown should have decreased dramatically during the quarantine.

So, we compared the outcomes using the normalized representation of both charts.

From what we can see, the growing trend is a bit more aggressive for the city center. Because of the strong initial fall, faster growth is something we should expect. As time went by, restrictions became weaker and more people returned to normal life visiting places located in the downtown.

If we take a closer look at June and July, we will see the trend (dotted line) is almost the same for the city center and Greater London despite the fluctuations in real numbers. This may mean that the walking patterns are getting back to normal at approximately the same pace in the center and the rest of the city.

A closer look at June and July: CCTV vs. Apple

As a conclusion of this Sofa Report, I would state that our approach certainly doesn’t work for bigger areas just because of the lack of video-surveillance. However, for local communities equipped with cameras, it’s quite comparable to the results of giants like Apple and Google.

All the computations here were done using the casual desktop CPUs and not-the-last-gen GPUs.

from Hacker News https://ift.tt/2WCUzLo