Artificial Intelligence: Deep Learning for ADAS/AD

Virtuelles Fahrzeug auf einer virtuellen Straße

The mobility of the future will be shaped by deep learning (DL) algorithms and neural networks. The methods of artificial intelligence (AI) are taking up an ever-increasing proportion of the software stack of automated driving functions and they form an essential building block of future autonomous vehicles, which will transform mobility and provide greater efficiency, more safety, more environmental friendliness, and lower costs.

For that reason, Bertrandt is working intensively on the topic of AI / deep learning in the field of ADAS / AD and provides development and validation services for solutions that are tailor-made to the customer. Bertrandt covers the entire deep learning workflow, from the creation of the data to the architecture development of neural networks and their training, as well as the embedded implementation of deep learning algorithms on the ECU (Electronic Control Unit) and the validation of neural networks for the safety of automated driving.

Learn more about our extensive know-how in the entire Deep Learning workflow:


Our range of services at a glance

Do you require high-quality data for training and would you like to save money and time? No problem! Our team of experts will help you to find the right data approach for your AI function and your special use case. Our specialists have wide-ranging experience in using the most up-to-date approaches for creating state-of-the-art synthetic data based on the latest research. If the ideal approach does not yet exist for your application, we will develop it together with you.  

  • Selection of the suitable type of data for your use case
  • Use of synthetic data for training/validating AI functions
  • Generation of various types of synthetic and simulated data (using a 3D engine, with copy & paste approaches, and generative neural networks)
  • Data generation with the aid of generative neural networks (GAN)

Are you looking for tailor-made deep learning algorithms for your application? We would be happy to develop them for you! Our experts closely follow the current state of research in order to implement the best possible solution for your challenge and to adapt the algorithms to your use case.

If the suitable algorithm is already available, we will take over the job of training and adaptation for your use case. If your application requires a tailor-made neural network or a combination of several approaches, we will develop advanced deep learning algorithms for optimum performance. If several sensors are involved, we have experience in the very latest deep learning-based methods for sensor fusion. And to ensure that the complex architecture of a neural network is able to run in the vehicle in real time on the ECU, our specialists will carry out comprehensive runtime optimization and adaptation to the target hardware.

Architecture development and training:

  • Architecture development of neural networks
  • Training of neural networks
  • AI/deep learning-based sensor data fusion of camera, lidar, and radar data
  • Knowledge of the latest research in environmental perception and motion prediction in the field of ADAS/AD

Implementation of deep learning algorithms on ECUs:

  • Embedded implementation of deep learning algorithms on high-performance ECUs (Electronic Control Units)
  • Runtime and resource optimization
  • Adaptation of the algorithm, computing power, and ECU for your use case

Are you sure that your AI algorithms work safely and reliably on the road? What happens if your camera is ‘blinded’ by glare or if a traffic sign is damaged?

The validation of AI algorithms is essential for the safety of automated vehicles, and this is something that Bertrandt works on very intensively. Neural networks function in a fundamentally different way than classic rule-based algorithms. For that reason, our experts are developing AI-specific methods, based on fixed criteria and statistical indicators, in order to validate the results of neural networks and their security against possible attacks. This validation includes the robustness and explainability of neural networks, the correctness of their results, and the completeness of the data sets used. What is more, it is expensive and time-consuming to acquire suitable data particularly for safety-critical scenarios that rarely occur. Bertrandt is therefore also working on the usability of synthetic and simulated data for AI validation.  

Validation of neural networks:

  • Selection of appropriate validation metrics depending on the application and use case
  • AI validation tool developed in-house with implemented validation metrics
  • Use of synthetic data for AI validation
  • Generation of critical scenarios / test data
  • Carrying out experiments for AI validation

Your Contact

Dr. Torsten Butz

Vice President Operations – Electronics & Virtual Testing Solutions

Related Topics