Artificial intelligence (AI) is a key technology for piloted driving – that’s why Audi and strong partners from the electronics industry are jointly developing game-changing know-how in the field of machine learning. At the Consumer Electronics Show (CES) in Las Vegas Audi presented the Audi Q7 deep learning concept, a piloted driving car made possible thanks to collaboration with NVIDIA.
In conjunction with the CES keynote address by NVIDIA, Audi is demonstrating the intelligence of the Q7 deep learning concept on a specially designed, variable open area for piloted driving. The car orients itself by means of a front camera with 2 megapixel resolution, and the camera communicates with an NVIDIA Drive PX 2 processing unit, which in turn controls the steering with high precision. The high-performance controller is specially engineered for piloted driving applications.
Serving as the core of the software are deep neural networks that experts from Audi and NVIDIA have trained specifically for autonomous driving and recognition of dynamic traffic control signals. Beginning with a human driver at the wheel, the Audi Q7 deep learning concept gained a limited familiarity with the route and the surroundings, by means of observation and with the help of additional training cameras. That established a correlation between the driver’s reactions and the occurrences detected by the cameras. So during the subsequent demonstration drives the car is able to understand instructions, like from a temporary traffic signal, interpret them right away and act as the situation requires. When a corresponding signal appears, the concept car immediately changes the driving strategy and selects either the short route or the long one. The design of the system is so robust that it can even cope with disturbance variables such as changing weather and light conditions. It masters its tasks day and night, and even in direct sunlight or harsh artificial light.
The learning methods used for the Audi Q7 deep learning concept are essentially very much like those of deep reinforcement learning. This method was the underlying principle behind the Audi presence at the Conference and Workshop on Neural Information Processing Systems (NIPS), an AI event held in Barcelona in December. There, the neural networks – which are similar to the human brain – were also trained for a particular application. While the 1:8 scale model car at NIPS learned how to park through trial and error, during the training runs the network of the Audi Q7 deep learning concept receives concrete data it finds relevant – in other words, it learns from the driver.
Another Audi key partner is Mobileye, whose image processing chip also is integrated in the zFAS. The high-tech Israeli company is the world leader in the field of image recognition for automotive applications. Mobileye is already supplying a camera for use in a range of Audi models – the Audi Q7, the A4/A5 series and the new Q5 – and the product’s image processing software can recognise a large number of objects. These include lane markings, vehicles, traffic signs and pedestrians. Today, defining the characteristics needed to clearly classify objects is still done manually.