High-Resolution 2D MIMO Radars for Traffic Gesture Recognition

Autonomous driving places high demands on auto-motive sensor suites in terms of environmental perception. This includes radar sensors, which must offer ever better imaging performance. As a consequence, imaging radars with large apertures and 2D beamforming capabilities with high-angular resolutions have emerged in recent years. On the algorithm side, these sensors open up new possibilities for deep learning-based environmental perception. Hence, this paper explores radar-based traffic gesture recognition with a 192-channel radar sensor with a large virtual aperture. The sparse array has an angular resolution of 1.45° and 6.17° in azimuth and elevation, respectively, which provides accurate and detailed radar point clouds. Subsequently, a convolutional neural network (CNN) for spectrogram-based classification and a PointNet-based model (PointNet+LSTM) for direct point cloud processing are trained on a traffic gesture dataset with 20 participants. PointNet+LSTM reaches a single-sensor accuracy of 90.8 % with pretraining from simulation data, demonstrating the great potential of radars with high angular resolution for reliable gesture recognition.