Santa Clara Transportation Dept.
When an autonomous Google Lexus came across two tiny sandbags on the road in California last month, it slowed down and stopped. Then, it pulled out around the bags,.
This accident, the first caused by a self-driving car after, suggests that Google’s vehicles still need more common sense. A human driver would probably either have just driven over the sandbags.
Researchers at the University of La Laguna in the Canary Islands are building a system which might have helped Google’s Lexus see the sandbags for the non-threat they were. They use Microsoft Kinect cameras, originally developed for the Xbox One gaming console, to improve self-driving cars’ obstacle avoidance at close range.
Self-driving cars typically use a combination of sensors to detect and avoid obstacles. Radar and laser-based lidar systems are generally used for objects at long range, while ultrasonic detectors and stereo cameras sense cars and pedestrians closer in.
Obstacles close to the ground like ramps, kerbs and sandbags are difficult to make out, says Javier Hernandez-Aceituno, lead author of the study: “Laser-based sensors are not suitable for this task because they detect ramps as obstacles. Ultrasonic sensors are also unsuitable due to their low precision.”
Hernandez-Aceituno decided to try using a Kinect, a depth-sensing camera that uses an infrared laser to capture an instantaneous 3D map of objects up to about 4 metres away.
Last year, he and his colleagues installed a Kinect on an experimental golf cart called Verdino. A low-speed self-driving vehicle, Verdino is also equipped with laser rangefinders and stereo cameras from a PlayStation 4.
They set the Verdino loose on an outdoor course with ramps, kerbs and stairs, using one obstacle detection program to process data from all the sensors.
The laser rangefinder ignored the lowest steps and incorrectly decided that the ramp was too steep to navigate. The camera gave inaccurate results for very near and far obstacles, and suffered from false detections.
The Kinect, however, produced more accurate results and fewer false positives than the stereo camera. Its biggest problem was spurious obstacles created by reflections of sunlight, although Hernandez-Aceituno was able to filter these out.
“The Kinect sensor vastly outperforms stereo vision at accurately detecting obstacles on the road,” says Hernandez-Aceituno, “[and] allows an autonomous vehicle to navigate safely in areas where laser rangefinders cannot detect obstacles.”
More on these topics:
This entry passed through the Full-Text RSS service – if this is your content and you’re reading it on someone else’s site, please read the FAQ at fivefilters.org/content-only/faq.php#publishers.