Side-by-side, the images only resemble one another in terms of geometrics.
In the left photograph, a lone rural road stretches out, edged by fields of grass. A stop sign stands on the road’s right side and a utility pole with electrical wires is center.
The right image is a splash of colors, the road represented by purple, the road markings in orange, the stop sign in pink. Each of the 12 colors represents a category of object one might encounter on the road.
One day, this may be how driverless cars see.
“Vision is our most powerful sense and driverless cars will also need to see,” said Prof. Robert Cipolla, of the Univ. of Cambridge. “But teaching a machine to see is far more difficult than it sounds.”
Cipolla and colleagues have designed two new systems that can recognize objects in the road and orient a user’s location with merely cameras, including smartphone cameras.
The first system—called SegNet—eschewed sensors such as radar and LIDAR to accomplish object recognition. Researchers trained the system using a form of machine learning known as supervised learning. A team of Cambridge undergrads labeled every pixel in 5,000 images, and trained the machine system with the information for two days.
The method was similar to how humans learn via example. However, with a machine it’s not a simple one and done task.
For instance, a child may successfully recognize what a bicycle is after one example, and then apply that knowledge to different types of bicycles. A machine may need many more—perhaps thousands of more—examples to achieve the same recognition ability.
“It’s remarkably good at recognizing things in an image, because it’s had so much practice,” said Alex Kendall, a PhD student in the the university’s Dept. of Engineering. “However, there are a million knobs that we can turn to fine-tune the system so that it keeps getting better.”
Trained primarily with images from the highway and urban environments, the system does surprisingly well in rural environments. It can be demoed here.
While SegNet detects a car’s surroundings, a second system is capable of determining a user’s location by analyzing a single image of the surrounding environment.
It can work indoors, in tunnels, and other areas where GPS systems are unreliable. The researchers tested the system on a kilometer-long stretch of the street King’s Parade in Cambridge. It successfully determined location and orientation within a few meters and a few degrees, according to the researchers.
While the advances are promising, it may be some time before they’re implemented into autonomous vehicles.
“In the short term, we’re more likely to see this sort of system on a domestic robot—such as a robotic vacuum cleaner, for instance,” said Cipolla. “It will take time before drivers can fully trust an autonomous car, but the more effective and accurate we make these technologies, the closer we are to the widespread adoption of driverless cars and other types of autonomous robotics.”