On the second day of the ATZ conference on automated driving, almost 20 presentations dealt with topics such as environment recognition, data and networking. Artificial intelligence is supposed to act as a support for the sensors.
Jörg Schrepfer, Head of Driving Assistance Research at Valeo, Germany
Uli Regenscheit
The participants of the panel discussion at the end of the first day of the event all agreed: Artificial intelligence (AI) will be part of autonomous driving in the future, and in some cases it already is. The first lecture of the second day also referred to this. Jörg Schrepfer, Head of Driving Assistance Research at Valeo, reported that it will be a combination of sensors and the right algorithms that will pave the way for autonomous driving.
Although artificial intelligence can act as a support, the appropriate set of sensors for data acquisition remains the basis for automated driving functions. Depending on the requirements, the sets of sensors can differ. There is no ideal solution that serves all OEMs equally well. Each sensor has advantages and disadvantages that are compensated by redundancy: Cameras do not deliver optimal results in bad weather conditions. A laser scanner on the other hand can capture most situations comparatively well. Although ultrasonic sensors can cover the proximity, radar also displays its strengths at longer distances.
Huge amounts of data have to be processed
The gigantic amounts of data generated that have to be processed pose major problems for the engineers. Raphael Pfeffer of IPG explained that although training data are decisive for the performance of an AI, the evaluation of the data can theoretically take years. Manual annotation, i.e. the marking of relevant information such as pedestrians, cyclists, cars, buses or traffic signs, takes around 3.5 minutes per image with today's technology. With a typical training dataset of 100,000 images, the annotation would take 737 days. Pfeffer did, however, admit that this was quite a bold example, because in practice not every single image has to be processed.
If the images are generated artificially with the help of simulation software, a lot of time can be saved. In this context it is interesting to note that the quality of the data that is based exclusively on simulation is not particularly good, but a combination with real driving data can significantly improve the quality. Even datasets containing five percent of real driving data in combination with 95 percent of simulation data provide very good results.
A combination of simulation and real driving data
Experts disagree on the amount of test kilometres needed to reliably validate systems. The figures range from 240 million to several billion kilometres. Simon Tiedemann of Elektrobit Automotive GmbH demonstrated that this generates enormous amounts of data that also have to be processed or evaluated. In the case of collecting real driving data, using 100 vehicles, 240 million kilometres would take 25 years. At the moment simulating this with 100 computers would take five years. Using a cloud, 240 million kilometres could be simulated in just two weeks - an enormous acceleration in terms of development, which does, however, cost a lot of money. In addition, the huge amounts of data that are in the petabyte range must first be uploaded to the cloud. Even with fast uploads this takes comparatively long.
Secondary activities may distract the driver
But it was not only the technical aspects that were discussed at the symposium. After all, customers must also accept and use the automated driving functions. In the session on users and acceptance, Paula Laßmann from the University of Stuttgart spoke about the balancing act between demanding too little and too much from the driver in semi-automated driving. Using test person studies in the simulator, Laßmann determined which secondary activities distract the driver to what extent and how their reaction times change. Real driving studies are planned for the coming year, but it is already foreseeable that visual secondary activities are rather unsuitable.
In further lectures, the participants informed themselves about sensor data fusion, algorithms for situation analysis, opportunities and risks of automotive ethernet or about plausibility as a prerequisite for trustworthy artificial intelligence.