Articles, Level 5 Autonomous Driving, Smart Mobility

Automotive Tech.AD Diary Part 1 - Keynotes by Nvidia & infoware

At the Automotive Tech.AD 2017 in Berlin the autonomous driving & automated reality community gathered to discuss the challenges of AI supported, cognitive cars in regards to complexity and penetration of in-vehicle infotainment systems and ADAS. The event was packed with inspiring keynotes and use cases – we put together the most exciting presentations and insights to provide a comprehensive review of Automotive Tech.AD Berlin. We take off with the presentations from Nvidia and infoware.

Nvidia: Artificial Intelligence & Deep Learning for realising Automated Driving

In the beginning of conference day 1 Sahin Kirtavit, Senior Director of Automotive Solutions from Nvidia, went into AI & Deep Learning for realizing automated driving. Nvidia is mainly known for its hardware in computer games, but recently the company started to focus more on other fields of activity. With artificial intelligence and visualization technologies based on their graphics processors, one can create virtual, augmented reality or a mix of them (MR – Mixed Reality). In addition, Nvidia is actively engaged in researching autonomous driving. Above all, the deep learning process is playing a crucial role for the development. It can be applied to data collection, processing and the output of the evaluated data back to the cars.

"Artificial Intelligence is the ultimate Challenge"

AI facilitates data processing and enables to work with larger models and higher computing capacities. Algorithms also improved which allows deeper insights. Even just considering image recognition Microsoft managed to achieve an increase by a factor of 16 during the last 3 years, stated Mr. Kirtavit.
While in 2012 AlexaNet still had 8 levels and was developing 1.4 GFLOP (Giga Floating Point Operations per Second) data with a 16% error probability, ResNet 2015 had 152 levels. The computing capacity was 22.6 GFLOP and the error rate was around 3% percent. Speech recognition performance, as an example Baidu was mentioned, increased tenfold within a year. In 2014 Deep Speech 1 still had a performance of 80 GFLOP and 8% error probability. A year later, with Deep Speech 2, 465 GFLOPs were processed at a 5% error rate.
The Deep Neural Networks (DNN) deliver data to the computer in the car. Data is synchronized with the HD map through the Nvidia PX2 AI car computing platform. Sensor data is sent the data center, processed with the Nvidia DGX1 in the training network and sent back to the vehicles. However, a new platform will have even faster results in the future: XAVIER is an AI Supercomputer SOC with 20 TOPS DL (Deep Learning Tera-Operations per Second) performance, but with only 20 watt power consumption - a quarter of what PX2 consumes. In addition, XAVIER is smaller, while the SPECint benchmark has increased from 120 to 160. This allows AI to adapt more quickly to the environment, which is particularly relevant for automating city traffic.

AI Sensor Technology - The Gamechanger

According to Sahin Kirtavit Nvidia wants to make greater use of the AI for sensor technology and for driving the cars, not only for mapping-related purposes. Therefore the company created an ecosystem of projects enabled through the Deep Neural Networks (DNN) mentioned above. Here Nvidia is partnering with HERE, TomTom, Zenrin and Baidu. The AI Co-Pilot can also recognize the driver, follows head and eye movements and can even read the lips with a 70% success rate. He even went a step further: why not shape the future society and health care with Artificial Intelligence? With that thought in mind Mr. Kirtavit closed his presentation.

infoware: Will Software be the future Differentiator of Cars?

Next up was Thomas Schulte-Hillen, CEO of infoware. Latter is dealing with navigation and vehicle automation. The key topics of his talk were data security and connectivity with regard to the automated car. As Mr. Schulte-Hillen is an avowed racing fan he pointed out the relevance of connectivity using the example of offroad racing. Connectivity allows drivers and teams to find out the perfect angle for a corner without seeing the track directly.

Following the Electronic Horizon

Getting an overview about what is coming up next, can be very convenient – not only for race drivers. The software solution “Electronic Horizon” provides information on what is lying ahead the next 2 kilometers. The technology for transmitting data via Cloud inspired infoware to develop the system. In order to find the best path and avoid traffic congestions or track-calmed sector, the user choose the “MMP”, the Most Probable Path. The number lanes and crossroads is taken into account as well as speed limits, traffic conditions and topography.

API for Data Transfer

The system’s architecture is scalable and also designed for automated driving. Mapping data from HERE or TomTom is fed into data streams with additional information that is transferred to the car. This calls for regularly updated software to keep the vehicles up to date. Currency is regulated by an API, a serial interface for data delivery. The relevance of certain data is varying, depending on the current region, weather, street conditions and the curve shapes the car is facing. Mr. Schulte-Hillen showed a map pinning all so called “Killer Curves” on it. The number of dangerous corners in Germany is surprisingly high – the Electronic Horizon could mitigate this source of danger.
In order to provide safety the system should be able to rely on precise data. Especially for speed limits the car needs to know the exact position of a traffic sign. The current problem is the time required to process the information via the cloud (Capturing, Publishing, Download, Computing and Distribution). The process can take up to 15 minutes, which is not close to real-time data at all. Many drivers would be stuck in traffic jams or get flashed for speeding without being warned about what is coming up. For this reason Schulte-Hillen sees software as a differentiating factor for vehicles which motivates infoware to work on a solution approach for the data processing in the API.
That is it for today, stay tuned for the next part of the Automotive Tech.AD Berlin Diary with more stunning speakers and presentations!