
Research project leans on AI to improve automotive radar
In the next two years, Astyx GmbH (project coordination), BIT Technology Solutions GmbH and the German Research Center for Artificial Intelligence (DFKI) will implement an automotive radar system for use in highly automated (Level 4) and autonomous driving (Level 5). The aim of the project is to significantly increase the quality of the measurement data of the continuous wave radar. With this radar technology, radar signals are emitted continuously during the measurement and reflections are measured.
The big advantage of radar-based sensing over camera-based methods or laser sensors (lidar) is the direct measurement of the object speed and the robustness against weather influences such as fog or snow. Disadvantages are possible errors in signal processing. These can be caused by speed ambiguities or multipath propagation. However, highly automated driving functions require very high accuracy and robustness. In the project AuRoRaS the physically caused disadvantages of radar sensors are to be recognized and eliminated by innovative AI methods.
In addition to its high-resolution radar equipment, Astyx also contributes its specialist knowledge in software-supported object recognition. This know-how in the areas of 3D object recognition from radar point clouds and deep learning-based object recognition will be used to improve the AI-based point cloud extraction from the radar raw data. Astyx is also developing synchronized data recording, geometric calibration of sensors and the development of data interfaces and tools for annotating real training and test data.
BIT Technology Solutions develops a synthetic, physically based simulation of the radar sensor data as well as the required reference data (Ground Truth). This simulation environment serves as a basis for a scalable and efficient training of the AI and the qualitative validation of the AI algorithms. The concept of BIT Technology enables research in the field of simulation of the entire spectrum of electromagnetic waves and the subsequent generation of corresponding synthetic data for training and validation. By combining simulated and real radar measurement data, the accuracy and robustness of the AI methods is significantly improved when cleaning up the measurement data and thus also the environment recognition of an autonomous vehicle based on it.
Based on its experience in the field of machine learning, the German Research Center for Artificial Intelligence (DFKI) is responsible for a large part of the improvement of data quality through newly researched signal processing steps using deep neural networks. These learning procedures require a large representative database, which otherwise could only be completely covered by real data in a very complex and costly way. The effect of multipath propagation in radar measurement is to be detected and compensated by machine learning methods. The measurement data of the radar sensor are analyzed with the aid of neural networks and incorrect values are subsequently removed.
