University’s self-driving car is as good as any automaker’s
The “Autonomoose” is a self-driving Lincoln sedan developed and undergoing real-world testing by a team of graduates and engineers at the University of Waterloo. What’s remarkable is the car’s capabilities are on par with the other fully automated rides researched and developed at the manufacturer level. Remember, this ’Moose is the product of a team building on a university-type budget, and not some deep-pocketed manufacturer laboratory. To put the cost of developing the fully autonomous car into perspective: the automotive and related industries have spent $8 billion on the technology in the past five years. It’s not the hardware, but rather the software that’s the complex part of the equation. The data cruncher has to be able to react to any given situation correctly, and it must do so each and every time. While the Lincoln MKZ Hybrid-based ’Moose is somewhat ungainly looking, it functions to a tee, and it did so on a very wet and rainy day. For some vehicles, inclement weather would be enough to postpone a demonstration. Not here, as ’Moose went out and completed its automated driving tasks without missing a beat. Sitting atop the roof are the cameras and the lidar (light detection and ranging) unit. There are eight cameras, which give a 360-degree view around the vehicle. The information generated by the “eyes” is key to autonomous driving. The problem is most of the software that looks at the camera-generated 2-D images will discard the information if the image is less than picture-perfect. In ’Moose’s case, as long as 70 to 75 per cent of the detail in the image is usable, the system continues to crunch the data. In this case, the raindrops on the camera’s lens were not enough to render the image’s information useless. The lidar unit stands proud in the centre of ’Moose’s roof. It scans the environment 10 times a second to create a second three-dimensional view of the surroundings. It not only detects trees, guardrails and other obstacles, it’s smart enough to learn its surroundings, so the next time the computer “sees” a familiar building it has an important point of reference. The battery of cameras and the lidar work in unison and develop a complex “map” that identifies everything in the immediate area and what lies further out. Another key part of the puzzle is the high-definition mapping. It’s designed so the system does not need painted lines to know where a lane lies or where the next stop sign sits; it is all contained in the data. When viewed on a screen, it shows three basic lines. There’s one for each side of the lane and a third that traces the middle of the lane, which is where the car needs to be. Three-dimensional dynamic object detection tracks other vehicles in real time. Once it latches onto a car, it tracks its progress until the car has passed the ’Moose. It also predicts the probable path of the car to ensure it is not moving into ’Moose’s lane. All of the information is fed to a deep neural network to determine the best course of action at any given time. Of course, there are myriad other sensors, including accelerometers and wheel-speed sensors, and it has a rule-based behaviour planner. The latter recognizes stop signs and the convention mandates it wait for three seconds at a stop sign before taking one final look around and making the decision to proceed through the intersection.