Vancouver's rainy weather could pose challenge for self-driving cars
Self-driving cars are coming, but operating them on dark, rainy days, like Metro Vancouver is experiencing this week, could pose some challenges.
That’s why computer vision researchers at UBC collected training data for autonomous vehicles, everything from pedestrians in dark clothing to objects that reflect light in the rain.
Jim Little, a computer vision expert and senior author of a study called Raincouver, says current training algorithms work well in good lighting and clear weather, but the data is not there yet for rainstorms or snow flurries.
“We have a hard enough time driving in the rain,” said Little. “But if we can do it, computers systems can.”
The Raincouver data, which was published in June, is the first scene parsing benchmark to focus on challenging rainy driving conditions, during the day, at dusk, and at night. Though Little noted that companies like Google and Uber are most likely doing their own similar research. The UBC data has been made available for free online.
For the study, lead author Fred Tung drove around in a rainstorm. Video sequences were collected using a digital camera mounted on the dashboard of his car. Then the researchers took each frame and drew lines around the objects or people.
They wanted to teach the computer how to recognize pedestrians, cyclists, and objects such as other cars and fire hydrants in poor visibility caused by rain. He admits it’s a massive challenge and far from being complete.
Many of the video sequences are captured at dusk and at night, when the vision system is challenged by severe glare from streetlights, traffic lights, and vehicle headlights. Oncoming vehicles with activated headlights often appear as a bright set of lights.
“The glare from headlights is a major problem,” said Tung, adding they hope with enough data the computer will be able to learn to identify glare and reflections off the road.
Tung said they label them as vehicles, and the scene parsing algorithm has to learn to make use of visual context to distinguish oncoming vehicles from other sources of light and glare.
Tung said although much more data is needed to ensure self-driving cars can handle the bad weather, he is optimistic that there will be enough in three to five years.
He hopes their data will be used in combination with other weather-detecting sensors for self-driving cars, for example LIDAR, which is a laser system to detect objects, and GPS.
via Vancouver Sun
