What’s the Difference between Zuckerberg and a Cardigan?

Computer vision and machine learning for intelligent autonomous systems are among the hottest—and most complex—technologies under development today. It is the holy grail that will allow autonomous vehicles detect and avoid obstacle and support general autonomous behavior and planning.

An image recognition engine is trained to classify general categories of objects. It recognizes vehicles, pedestrian and cyclists, as well as inanimate objects. Based on a priori knowledge of each category’s behavior pattern and range of movements, the car could predict what actions to expect and respond appropriately.

The ability to classify objects fast enough to ensure safety in normal traffic speed continues to be a challenge. Recognizing traffic lane markers and traffic lights is far less complex than recognizing and responding to the tacit complex interaction between drivers, cyclists and pedestrians.

For instance, recognizing the left turn signal of car and slowing down or yielding, as necessary, is a reasonable expectation. In fact, when all cars are equipped with vehicle-to-vehicle communication systems, turn signals will be largely superfluous, at least as far as the autonomous cars are concerned. But do we also expect a driverless car to notice that a cyclist is extending her left arm and anticipate she will turn left? What about her right arm? Or is it left arm extended in a 90 degrees angle that signals a right turn?

It is a difficult task. Volvo proposed using a smart helmet to connect drivers and cyclists. But what about drivers? Pedestrians?

For some light reading on the subject, turn to a recent Bloomberg Business article. As it turned out, some artificial intelligence computer vision systems mistake Zuckerberg for a cardigan, but, oddly enough, they do an excellent job recognizing cats…  Click for a larger image that compares the performance of top AI based computer vision systems and a good illustration of the  the complexity and state of the art of image recognition.