PORTLAND: Ford-financed researchers conclude that you do not need fancy radars or expensive lasers to successfully build self-driving cars. According to these researchers at the University of Michigan (Ann Arbor), all you need is a single inexpensive camera, access to 3-D maps and the University of Michigan’s custom software suite for self-driving Ford Fusions.
“We adapted off the shelf equipment,” University of Michigan professor, Ryan Eustice, told EE Times. “The algorithm makes use of a standard camera and a consumer-grade GPU for synthesizing the synthetic images for self-driving from remade 3D maps.”
So far, Eustice, in collaboration with doctoral candidate Ryan Wolcott, are spending most of their time perfecting the algorithms, which eliminate shadows, identify road edges and recognize dynamically moving obstacles. For now a lidar is used to generate the 3D scenes, but once made, the self-driving car needs only a cheap video camera to enable ultra-inexpensive self-driving cars.
“Once the lidar reference map is derived, only a single forward looking monocular camera is needed to use the localization framework,” Eustice told us. “Hence enabling low-cost applications of this technology.”
The researchers claim the technique gives them centimeter accuracy for up to $10,000 less than the expensive 3D laser scanning technologies used by others researching self-driving cars such as Google. The key is to create a real-time self-driving map by making thousands of comparisons per second between a survey 3-D map stored in the car’s system or streamed over the cellular network.
Self-driving was accomplished by comparing the survey 3D images created by their algorithms with the real-time images streaming in from a single camera. Their biggest engineering challenge was to process the real-time data with an off-the-shelf GPU fast enough to have it ready for comparison with the real-time image stream.
The team has successfully tested the system on streets in downtown Ann Arbor — with a driver at the wheel ready to take over at any instant. A special mini-city test facility is currently being constructed specifically for testing self-driving cars without the need for any driver at all. Eventually they want to eliminate the need to use any mapping information, so that vehicles can travel anywhere.
“We continue to work on all aspects of automated driving with a goal of better than human-level performance,” Eustice told us.
The biggest obstacles to overcome are weather conditions, light levels and unexpected road obstacles, such as double-parked cars and pylons marking roadwork.
Funding was provided by Ford Motor Co. via the Ford-UM Alliance. Wolcott was personally supported by the SMART Scholarship for Service Program by the U.S. Department of Defense.
