With long curly locks, sandy moustache, and driving cap, Edward Zimmerman looks like the fiddler in a Scottish folk band. To say he looks out of place as a presenter at the GPU Technology Conference is even something of an understatement—Zimmerman would look out of place behind a computer in a public library.
His presentation is likewise out of place. For one thing, its focus is on public transportation infrastructure, a domain that’s not likely to get anyone rich and-or famous. Specifically, Zimmerman, as a cyclist, wants to fix urban bicycling—or at least make it better and perhaps safer. He wants to do this by optimizing traffic lights for bicycle traffic using deep learning.
Cyclists and drivers are aware of the phenomenon of the “green wave” even if they don’t know it by name. It’s where you get the timing of traffic lights on a city street just perfect and are able ride a “wave” of green traffic lights through the city. The wave sensation occurs because the lights are changing from red to green in front of you (and green to red behind you), and you’re able to watch the transition unfurl at a pace roughly matching your own. It’s wicked satisfying.
A green wave usually isn’t an accident: traffic engineers try to time lights in concert with average speeds along a roadway. For auto traffic, this can mean less noise, better fuel efficiency, and less pollution. For a cyclist, meanwhile, stopping and starting at every light is more than an interruption, it’s a lot of added physical effort.
In some cases, engineers are taking a more proactive approach to green waves by adding sensors into the system that can register cars and-or bikes and dynamically generate favorable light patterns. This is where Zimmerman—a cyclist himself—comes in.
Sensor-based detection systems are expensive because they tend to involve digging holes in the road for magnetometers, which often don’t pick up modern bicycle materials (like carbon-fiber) anyway. Zimmerman’s idea is to use traffic light-mounted cameras coupled with deep learning algorithms to detect cyclists as they approach intersections. Technically, it’s then an object-recognition problem, albeit one that has to work in a variety of conditions and settings.
Fortunately, being able to discriminate among different classes of object in variable or noisy conditions is pretty much what machine learning exists for. The catch is that it’s very computationally expensive. Picking out a cyclist from an image using a Raspberry Pi, Zimmerman says, is a 12 minute ordeal. That’s not useful.
Fortunately, it’s now possible to pack high-performance GPUs into embedded computers (small computers that interact in real-time with an environment, generally). Nvidia offers such as system in its Jetson development boards and this is what Zimmerman ultimately used to perfect his system. It was demonstrated in Germany last year. “We showed it pretty much worked,” he says.
Are GPU-enabled traffic lights an inevitability? Zimmerman’s work is temporarily on hold owing to his research partner’s illness, but he hopes to be back at it soon.