Two New System Can Help Driverless Cars See Better

Self-driving cars, in order to perform their seemingly magical feats of technological marvel, need to be able to answer three critical questions: where the car is, what's around the car, and what should the car do next based on the first two questions. That third question can be sufficiently answered by complex algorithms and software. Two different but related researches at the University of Cambridge are seeking to answer the other two.

"Where am I" seems like a question best suited for a GPS and, in most cases, it is. However, there are times when GPS is either inaccurate, slow, or even completely unavailable. A new system promises to supplement that using computer vision. By using a single color image, a computer is able to determine its precise location, within a few meters, sometimes even a few degrees, trumping the accuracy of most GPS systems. Using geometry, it can also accurately determine which side of the building it is facing, even if all sides look identical.

The second system, which has a more proper name of SegNet, also applies computer vision to answer the second question: "what's around me?". Here, the system analyzes an image of a scene it has never seen before and classify and color code objects into roads, street signs, buildings, pedestrians, cyclists, and more. And it doesn't even need expensive equipment to like radar or LIDAR to do so. It can just use the images coming from a camera or a smartphone. All in real-time speed.

The way SegNet learned how to do this is just as interesting as its ability. SegNet was actually "taught" by researchers how to identify objects by meticulously labeled every pixel in 5,000 images, with each image taking up 30 minutes to process. After teaching, SegNet was further trained to identify objects before it was put to the real world test. Almost like how a human also learns things through repetition, only multiplied a hundred, even a thousand, times over.

SegNet isn't just some theoretical system either. The system is available online for curious testers to upload photos and watch the magic of computer vision happen. It's uncertain, however, if car manufacturers will be willing to give up the millions they've invested in more sophisticated equipment, but backup systems such as these could prove to be handy.

VIA: EurekaAlert!