Researchers Develop Joint Reasoning Algorithm for Driverless Cars
Researchers from the University of Toronto have teamed up with Uber's Advanced Technologies Group to create an algorithm for autonomous cars. Click here to learn more.
Join the DZone community and get the full member experience.Join For Free
One of the main challenges is that autonomous vehicles are trained to detect, track, and predict objects independently from one another, which creates a significant amount of risk, should any individual task fail. A team from the University of Toronto’s department of computer science and Uber’s Advanced Technologies Group (ATG) in Toronto have developed an algorithm that allows the tasks to be brought together, all in as little as 30 milliseconds per frame.
“We try to optimize, as a whole, so we can correct mistakes between each of the tasks themselves,” the researchers explain. “When done jointly, uncertainty can be propagated and computation shared.”
The work, which was documented in a recently published paper, began with the collection of a huge dataset consisting of Li-DAR collected data from several North American cities. This data was then transformed into a point-cloud in a three-dimensional space so that it could be analyzed by their AI technology. That data is inherently unstructured, which makes it harder than most AI tasks.
“If the task is detecting objects, you can try to detect objects everywhere, but there’s too much free space, so a lot of computation is done for nothing. In a bird’s eye view, the objects we try to recognize sit on the ground, and, thus, it’s very efficient to reason about where things are,” the researchers say.
The AI tool developed by the team was built specifically to manage this kind of unstructured data.
“A picture is a 2-D grid. A 3-D model is a bunch of 3-D meshes. But here, what we capture [with Li-DAR] is just a bunch of points, and they are scattered in that space, which for traditional AI is very difficult to deal with,” they say.
AI is exceptionally good at working with images, largely because they’re usually common shapes, often rectangular, with the pixels themselves also rectangular. LiDAR data, however, is much more irregular, which presents a challenge for AI researchers.
Suffice to say, the team believes that their work can be applied to any unstructured data set, but it’s likely to be most commonly used in developing autonomous vehicle technology.
The team also wanted to increase the speed with which these computations were performed, and they did this via the creation of SBNet: Sparse Blocks Network for Fast Inference.
“We want the network to be as fast as possible so that it can detect and make decisions in real time based on the current situation,” they explain. “For example, humans look at certain regions we feel are important to perceive, so we apply this to self-driving.”
The algorithm they developed utilizes sparse computation based purely upon what regions were deemed important. This made the computation roughly ten times faster than existing methods. In other words, it allows the car to see everything, but it only computes what is regarded as important, thus, making it more efficient.
“So, when there’s a lot of cars [on the road], the computation doesn’t become too sparse, so we don’t miss any vehicles. But, when it’s sparse, it will adaptively change the computation,” the researchers explain.
The code for the algorithm has been freely released, as the team is confident it has a wide variety of possible use cases. They’re hopeful that the work will help to solve some core problems, both in autonomous vehicles and further afield, and will, therefore, extend its usefulness beyond mere academic curiosity. It will be interesting to see just how accurate that prediction is in the coming months.
Published at DZone with permission of Adi Gaskell, DZone MVB. See the original article here.
Opinions expressed by DZone contributors are their own.