Simulation enables training of deep learning algorithms in virtual driving tests

August 21, 2018 //By Christoph Hammerschmidt
Simulation enables training of deep learning algorithms in virtual driving tests
Highly automated driving requires a large number of tests to safeguard functions that can only partially be covered in real driving tests. This is why virtual test drives are gaining in importance, especially with regard to safeguarding automated driving functions and also through the use of algorithms based on artificial intelligence (AI). For the necessary training of such AI algorithms, the simulation software CarMaker can generate numerous data and create reproducible scenarios to secure automated functions. A particular advantage: these virtual test drives can be carried out in all development phases of the car.

Artificial intelligence is already used in many different fields of the automotive industry as a term for the mechanical imitation of thought and learning structures in the human brain: For example, to make decisions on driver assistance systems or for object classification and interpretation, powered by the sensors installed on the vehicle. Real driving videos with appropriate metadata are made available to the system during the learning process so that a vehicle can really drive independently and automatically with the aid of AI algorithms. Different road markings, road users, buildings, parked cars, etc. with various visibility and weather conditions such as fog, rain, day or night complete the scenarios, so that appropriate reactions can be learned. Neural networks trained in this way can be used in the vehicle in real time to identify other road users or traffic signs, for example, and to trigger suitable actions even in complex traffic situations.

With the simulation solution CarMaker from the software company IPG Automotive, car electronics developers can create reproducible data virtually and automatically label them directly with 100 percent accuracy. With real data this is only possible with enormous additional effort. Different scenarios, object lists for decision or route planning algorithms or automatically labelled video data for object recognition algorithms can be used to train the neural networks. The huge amount of real test drives actually required for this is thus minimized through the use of virtual journeys; time and costs are reduced.

The software also makes it possible to integrate and test AI algorithms in different stages and characteristics in scenarios - over the entire development period. "The use of CarMaker for virtual testing of the fully trained algorithms in real scenarios and in the context of the entire vehicle enables a higher degree of maturity in the application of the algorithm in the real vehicle. If the software solution is also used on high-performance computing clusters, larger test catalogs can be covered simultaneously, automatically and with subsequent evaluation and test report generation," explains Dominik Dörr, Business Development Manager Driver Assistance and Automated Driving at IPG Automotive.

Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.