“Inspiring Technology: 34 Breakthroughs”: Download IEEE’s 140th anniversary book for FREE.

Close bar

San Diego’s Connected Streetlights Learn to Spot Bicycles

The only catch is that they also count the ones hanging from bike racks and tossed into the beds of pickup trucks

2 min read
Image of a bike with a square around it next to cars on the street.

Why might a city planner want to know how many bicycles are on the streets, how fast they’re going, and in what direction they’re moving?

There are lots of reasons why, says Austin Ashe, general manager for intelligent cities at Current, a General Electric subsidiary. “By time stamping and aggregating the data, we can find out what parts of a city are used most and least by bicyclists, and [consider] what the city can do to enhance ridership or make existing areas that riders use more safe.”

“You can also see how this data could be used by bike share applications,” he said, “or integrated into traffic signals, for example, giving bicycles a protected left turn.”

Current designed and installed the City of San Diego’s 3,200 smart streetlights, each of which monitors roughly 36 by 54 meters of pavement. The network, which became operational in 2018, initially tracked only cars, both on the move and in parking spaces, and started using that data to place stop signs and time traffic signals.

San Diego then added data about pedestrian movements into the mix. And, in the second quarter of this year, the city updated its network to count bicycles and record their movements as well.

Developers used machine learning, Ashe said, to teach the system to “distinguish a bike from a car, a person, or a trash can.” It wasn’t the simplest of projects, he indicated, with “a lot of edge cases,” and complicated by the position of the sensors way up on lampposts—bicycles are far more distinctive when viewed from the side than from above.

Standard image analytics, Ashe said, “didn’t let us count ‘cyclists,’ rather we had to count what the classifier can detect, which is one bicycle and one person. Then we further analyzed these two data sets into an ‘insight,’ inferring that when a bicycle and a pedestrian are reported with the same location and the same time stamp, then they are some percentage likely to be a cyclist.”

Once the system is trained, the actual processing of what is and isn’t a bike takes place “at the edge”—meaning, in the streetlights themselves.

The developers who worked on the bike-counting software discovered a surprising glitch during real-world testing: the system was counting some bicycles that were actually not being ridden—for example, those hanging from bike racks on public buses, or tossed into the beds of pickup trucks.

The software has the same problem when counting pedestrians, Ashe said. “If a car drives by that is a convertible, we are counting it as both a car and a pedestrian, even though the ‘pedestrian’ is the driver.”

“Our next challenge is to fix that,” he said.

After the developers perfect the sensor network’s ability to count bicycles, Ashe isn’t sure what will come next. Perhaps, he says, the streetlights will start counting scooters, or shopping carts, or wheelchairs—data on any of these items would be useful from a city planner’s view.

The Conversation (0)