Teaching autonomous cars to drive in Canada

If you’ve ever wondered how the machine learning algorithms within autonomous vehicles learn how to drive, its largely by observation. Before they’re allowed on the road, such AI systems have logged countless hours in simulated driving environments constructed from training datasets (i.e. gigabytes of digital video and sensor data collected by sensor laden cars). Using that data, developers construct the virtual environment in which AI systems learn how to identify buildings, other vehicles, pedestrians, road signs, where the boundaries of the road are, etc.

Obviously, these training data sets are invaluable to researchers and autonomous vehicle developers, but they have, until now, largely been collected and constructed under sunny day driving conditions. While that may be great for self-driving cars in southern California, it does leave places like Canada, with its snowy winters, out of the loop or at a disadvantage.

According to the researchers at the University of Toronto and the University of Waterloo, self-driving systems taught to drive using such “blue sky” datasets often misidentify or are oblivious to pedestrians and other vehicles when tasked with navigating snowy roads or other less than ideal conditions.

To remedy that, the U of T and UWaterloo researchers have released the Canadian Adverse Driving Conditions (CADC) dataset, a pool of high quality sensor data designed to capture the realities of driving on snowy roads and in the less than ideal visibility of a typical Canadian winter.

Built over the past two winters, in and around Southern Ontario, the CADC dataset was collected using UWaterloo’s Autonomoose, a Lincoln MKZ hybrid equipped with GPS, a lidar scanner and eight onboard cameras, that can capture 10 images per second.

Over the past two winters, the teams drove the Autonomoose around southwestern Ontario clocking more than 1,000 kilometres of data, approximately 33 kilometres of which is in snowy conditions. Data in hand, the teams worked with California-based AI firm, Scale AI, to catalogue and tag approximately 178,000 passing vehicles and 83,000 pedestrians captured in the CADC data.

According to the researcher teams, the resulting validated and formatted data is now available to any autonomous vehicle researchers who want to use it, through a free-to-use, open source model, although commercial users are required to license the software. In addition, the teams have posted documentation and support tools on GitHub, along with a scientific article on arXiv.

“We’re hoping that both industry and academia go nuts with it,” says Steven Waslander, University of Toronto Institute for Aerospace Studies professor and founder of the Toronto Robotics and Artificial Intelligence Laboratory (TRAILab). “We want the world to be working on driving everywhere, and bad weather is a condition that is going to happen. We don’t want Canada to be 10 or 15 years behind simply because conditions can be a bit tougher up here.”
http://cadcd.uwaterloo.ca
www.utias.utoronto.ca