Uncrewed Systems Technology 043 l Auve Tech Iseauto taxi l Charging focus l Advanced Navigation Hydrus l UGVs insight l MVVS 116 l Windracers ULTRA l CES 2022 show report l ECUs focus I Distant Imagery

8 Platform one Researchers in Austria have used neural networks to improve the accuracy and reliability of radar sensors in autonomous vehicles (writes Nick Flaherty). The researchers, from the Technical University of Graz, and Infineon Technologies in the Robust and ExPlainable AI for Radarsensors project, are optimising their development over the next 3 years to integrate them with radar sensors. “The better the denoising [noise reduction] of interfering signals works, the more reliably the position and speed of objects can be determined,” said Professor Franz Pernkopf from the Institute of Signal Processing and Speech Communication at TU Graz. His research has shown that an AI system based on neural networks mitigates mutual interference in radar signals far better than the current state of the art. The researchers now want to optimise this AI model so that it also works outside learned patterns and recognises objects even more reliably. The researchers first developed model architectures for automatic noise suppression based on convolutional neural networks (CNNs). The TU Graz team trained a variety of these networks with noisy data and desired output values. In experiments, they identified particularly small and fast model architectures by analysing the memory space and the number of computing operations required per denoising process. The most efficient models were then compressed again by reducing the bit widths – that is, the number of bits used to store the model parameters. The result was an AI model with high filter performance and low energy consumption at the same time. The denoising results, with an F1 score (a measure of the accuracy of a test) of 89%, are almost equivalent to an object detection rate of undisturbed radar signals. This effectively removes the interfering signals completely from the measurement signal. With a bit width of 8 bits, the model achieves the same performance as comparable models with a bit width of 32 bits, but requires only 218 kbytes of memory. This corresponds to a reduction in storage space of 75%. “For our successful tests, we used interfering signals similar to those used for the training. We now want to improve the model so that it still works when the input signal deviates significantly from learned patterns,” said Prof Pernkopf. This would make radar sensors many times more robust with respect to interference from the environment. “Until now, even the smallest changes in the measurement data were enough for the output to collapse, and objects not to be detected or to be detected incorrectly, something which would be devastating in the autonomous driving use-case,” Prof Pernkopf said. The system has to cope with such challenges and notice when its own predictions are uncertain. Then, for example, it could respond with a secured emergency routine. To that end, the researchers want to find out how the system determines predictions and which influencing factors are decisive for this. For this purpose, the complicated model architecture is transferred into a linear model and simplified to make it visible. “We want to make the CNNs’ behaviour a bit more explainable,” Prof Pernkopf said. “We are not only interested in the output result, but also in its range of variation. The smaller the variance, the more certain the network is.” Sensors AI cuts interference April/May 2022 | Unmanned Systems Technology The REPAIR project is aiming to use machine learning to improve the quality of data from a radar sensor