RADRadar: Range-Angle-Doppler Feature Cube Reconstruction for Radar Scene Perception

Compared to cameras and lidars, millimeter-wave radar's capability to operate under all-weather, all-day conditions makes it indispensable in autonomous driving. The radar cube, comprising range (R), azimuth (A), and Doppler velocity (D) dimensions, serves as an effective data representation due to its well-structured format and rich information. However, the vast amount of information comes with high computational overhead, especially for the design of radar-based perception networks. How to efficiently represent such dense RAD tensors still remains an open challenge. To address this challenge, we propose a novel multi-view fusion architecture that decreases the need for computationally intensive operations on high-dimensional RAD data. To mitigate the information loss during view compression within multi-view architectures, we leverage complementary relationships between different radar views to efficiently construct a latent RAD representation via an outer product module. Before that, an attention-inspired feature similarity learning module is also introduced to ensure the efficacy of the outer product operation. In downstream perception tasks, our method, RADRadar, outperforms state-of-the-art methods on semantic segmentation on the public CARRADA and RADIal datasets. In terms of object detection on the RADIal dataset, RADRadar also demonstrates competitive performance, especially in the hard split with high radar perturbation.