Comparison of Neural Network Architectures for the Detection of Clutter in Automotive Radar Data

The unique properties of radar sensors, such as their robustness to bad weather conditions, make 'them an important part of the environment perception system of autonomous vehicles, However, the detection point clouds generated by automotive radar sensors contain a large amount of (clutter. These erroneous detections, which do not match any real object, are highly problematic for subsequent processing steps. To address this issue, we present three different neural network architectures for identifying such clutter detections. We compare their accuracy and speed using a large and diverse public data set In particular, a state-of-the-art PointNet++, a newly designed Transformer architecture and an optimized convolutional neural network (CNN) setup are examined, For the latter, the learned behavior and the influence of the chosen cell size of ‘input grid maps are analyzed in detail.