Enabling Azure @ Audi technology partner EFS uses deep learning to analyze roads for self-driving vehicles

Based in Gaimersheim, Germany, EFS is the number one partner of Audi in chassis development. It examines and helps implement future-looking technologies, including automated driving. As part of its research efforts, the company used Azure NC-series virtual machines powered by NVIDIA Tesla P100 GPUs to drive a deep learning AI solution that analyzes high-resolution two-dimensional images of roads. The purpose is to give self-driving vehicles a better understanding of those roads. EFS proved that the concept works, and the company can now move ahead with product development.

Imagine you’re in your car, about to head home after leaving the grocery store. A quick glance around reveals shopping carts, traffic signs, and other vehicles and shoppers. Your brain processes all this information and you decide when to move forward, what path to take, and how fast you should go. And before you know it, you’re on the road again. But suppose your car could do all of that for you.

Despite the inevitable ubiquity of self-driving cars, a fully autonomous point-to-point vehicle is still a work in progress. One of the technologies that’s likely to play a significant role in the research effort is deep learning—a subset of artificial intelligence (AI) and an extension of artificial neural networks. Because of its state-of-the-art performance, deep learning offers big data predictive analytics that automated driving systems benefit from.

Helping self-driving vehicles to think ahead

Elektronische Fahrwerksysteme GmbH (EFS) is testing deep learning technology for use in Audi vehicles. Founded in 2009, EFS is a joint venture of the GIGATRONIK Group and Audi Electronics Venture. “We’re examining use cases to see how we can apply deep learning to protect drivers and passengers,“ says Max Jesch, Software Developer at EFS. “One of those cases is to give the vehicle a way to gain a good understanding of the road around it.” That means the car has to conceptualize factors, such as relative distances, what areas are part of the road, and what objects are hazards versus ones drivers can safely ignore.

Jesch says that a lot of the things we think we know about driving are forms of anticipation. “You can’t see the road in some places; you just know that it’s there and can infer what to do next,” he explains. “Whether the road is curved, covered by leaves, or blocked by something that limits sight distance, it’s easy for you to figure it out, but not easy for a computer.”

Innovating with the right deep learning technologies

As part of its research to address these challenges, EFS is creating a solution powered by Microsoft Azure that analyzes pictures taken from a car’s front-facing high-resolution 2D camera. Jesch explains that EFS has never applied deep learning to processing images of this type, so the company needed a successful proof of concept before it could proceed with any product development.

That’s where EFS demonstrates its innovation. Says Jesch, “We created proprietary recursive algorithms to analyze images and took advantage of the massive capacity and scalability of Azure virtual machines to handle the computational load and data storage.” Specifically, EFS is using Azure NC-series VMs powered by NVIDIA Tesla P100 GPUs along with Azure Blob storage and Azure Disk storage. The Azure N-series is a family of Azure virtual machines (VMs) with GPU capabilities. The architecture of the NVIDIA GPU makes it well-suited to deep learning tasks. Regarding the training data, Jesch adds, “We’re using Blob storage for several terabytes’ worth of data; nothing fancy, but the simplicity of the storage solution makes our job a lot easier.”

The main reasons EFS chose to use the cloud are flexibility, availability, and storage integration. Says Jesch, “With Azure, we can have as much compute power as we need, whenever we need it. Functionality is available from different workspaces, which is great for cooperative working, and we get reliable integration of storage and data management technology.”

Gaming the learning system

EFS programmed what is effectively a video game that generates the deep learning training data (images). That way, the time-consuming work of labeling real-world data is reduced to a minimum. To handle the high-resolution images in a deep learning system, a progressively lower image resolution is used in each recursive step in the analysis. Processing an image with resolution that’s lower than the original takes less time.

Jesch notes that competing cloud providers offer services comparable to Azure; however, “We ultimately based our decision to choose Azure on the relationship we have with Microsoft and with the people who represent it and support us. The cooperation we got on this project has been awesome.”

The deep learning tests using Azure virtual machines were successful, and EFS sees the project as the precursor of breakthroughs to come. “The innovative ideas we’ve implemented so far give us trust in a new deep learning architecture and in solutions that will rely on it.” He concludes, “We still have a long way to go, but the project’s main takeaway is that as part of our autonomous driving research for Audi, we proved that it’s possible to use deep learning to analyze roads. That is a really big deal. As far as we know, EFS is the first company to do it on such a large scale.”

Find out more about EFS on Facebook and LinkedIn.

We’re using Blob storage for several terabytes’ worth of data; nothing fancy, but the simplicity of the storage solution makes our job a lot easier. 

Max Jesch: Software Developer

EFS

https://azure.microsoft.com/en-us/