The barrier to market for Google’s self-driving cars just lowered a smidge. Google research scientist Anelia Angelova unveiled at the IEEE International Conference on Robotics and Automation (ICRA) in Seattle on May 27 new technology that could cut costs on one of the more expensive components of self-driving vehicles.
The company presented research indicating that new pedestrian detection systems may be possible to implement without the use of spinning light detection and ranging (LIDAR) devices, which can often cost more than $10,000. Instead, cheap video cameras may be used for the same purpose of recognizing and tracking nearby pedestrians.
By using deep neural networks and advanced artificial intelligence, researchers reported that cameras may be able to outperform humans in object recognition. Such systems can outperform humans today when it comes to facial recognition, but object detection in a pedestrian setting is very slow. The camera must divide a view of the street into 100,000 tiny images and analyze each frame one by one.
A solution to this problem, the researcher presented, is a high-speed analysis that ignores those regions of the image where pedestrians are unlikely to be, resulting in a 60 to 100 percent increase in image recognition speed. In testing, Google’s system was able to accurately identify pedestrians in about a quarter second, Angelova said. While it’s progress, a time of .07 seconds is needed for practical use.
With more development, however, LIDAR could eventually disappear from self-driving technology, decreasing costs and bringing the technology closer to the mainstream.