Google Street View data can predict how you vote, researchers say

Google Street View can be incredibly useful for finding your way around a new city.

But a new project out of Stanford University shows that Google Street View is full of surprisingly useful demographic information, too.

It may even be able to predict who you'll vote for.

CBC technology columnist Dan Misener has the story.

How can Google Street View predict who I'm going to vote for?

For the past 10 years or so, Google has been driving special cars all over the world, taking panoramic photos of public streets.

• Google Street View team hits the ice to capture Arctic wildlife
• Sable Island now on Google Street View
• Winnipeg man uses Google Maps to successfully fight photo radar ticket

This means Google owns a huge database of street-level images that cover much of the world, including most of the U.S. and Canada.

We're talking about this today because a team of computer scientists managed to use Google Street View data to predict election results.

They did this by training an algorithm to scan through Google Street View, look for vehicles on the road and identify their makes and models.

For instance, by comparing the number of pickup trucks versus sedans in a given U.S. neighbourhood, they found they could accurately predict whether an electoral precinct would vote Democrat or Republican.

They shared their results in a paper, which is currently under peer review.

Beyond political trends, what else can be measured with Google Street View?

From a tech perspective, one of the most impressive parts of this project is how scientists trained a computer to reliably tell the difference between very similar cars.

For instance, their algorithm could distinguish between a 2007 Honda Accord and a 2008 Honda Accord, which are visually very similar, except for a subtle difference in tail lights.

It took the algorithm about two weeks to classify all the vehicles in 50 million images. By comparison, researchers say it would take a human being more than 15 years to do the same task.

By identifying the make, model, body type and year of each vehicle, researchers could predict not only the political trends of a neighbourhood, but demographics such as income, race and education level.

The type of car you drive offers all sorts of clues about who you are. Of course, this doesn't work well at the individual level but it's surprisingly accurate neighbourhood-by-neighbourhood.

How else could this vehicle-recognition technology by used?

The ability for computers to reliably identify cars and distinguish between different makes and models is relatively new. But we've already seen some interesting applications.

Last summer, I did a story about a company in the UK that uses vehicle-recognition technology to deliver customized billboard ads.

They installed a 12-metre-wide billboard at a busy roundabout in London, equipped with cameras that point back at traffic.

The cameras can identify your car, and tell its colour, age and whether it's gas, diesel or electric.

The billboard then shows targeted advertising, based on who's stopped at the lights.

So you might get an ad that says, "Hey you in the silver hatchback..." — that kind of thing.

The team at Stanford isn't focused on advertising, though. The researchers imagine this vehicle-recognition technology as a way to track the changing demographics and socioeconomic trends of a city or neighbourhood.

What are the limitations of this approach?

One obvious limitation is that this approach uses vehicles as a proxy for human beings.

Of course, not everyone owns a car. Here in Canada, about 78 per cent of the driving-age population owns a car.

Another limitation is that it only counts vehicles that you can see on public streets.

Google Street View covers most of the U.S. and Canada, but there are plenty of spots where their cameras don't reach. For instance, underground parking for people who live in condos.

Unless your car is out on the street when a Google Street View drives by, it won't be included.

Also, the car classification system used here only goes back to 1990. So if you're driving a car built before then, it won't be in the database.

The system isn't perfect, of course, which isn't to say it's not useful. But I think anytime we talk about big data, it's worth considering its biases, especially who gets left out.

What's next for this project?

The original idea was that Google Street View data could be an alternative to a traditional door-to-door census. Manually collecting data can be costly, labour-intensive and it can take a long time to complete.

The researchers hope that by using data that already exists and is updated more frequently than a census, you could monitor demographic and socioeconomic changes in a more up-to-date way at a fraction of the cost.

In the future, the researchers expect that self-driving cars could provide a huge amount of useful information, because most self-driving cars are equipped with cameras that constantly photograph the world around them.

The potential for real-time data collection is pretty incredible.

But of course, it also raises some pretty serious ethical concerns. Do I really want my car to be constantly recording, analyzing and estimating the income, education level and political views of everyone around me?