DeeperRadar is a radar-centric framework designed to co-design radar sensing and multi-modal 3D detection for autonomous mobility by learning a sparse acquisition pattern end-to-end within a fusion model. The core of this framework is a learnable MIMO design module, trained end-to-end in a fusion network that processes raw radar ADC data alongside camera images and LiDAR point clouds.
During training, the design module is supervised by other sensors, enabling the system to learn which receiver antennas to activate and how many to use effectively. At deployment, the design module is removed and replaced with the learned sparse subsampling mask, keeping the downstream model architecture unchanged.
Evaluated on the RADIal dataset, DeeperRadar identifies sparse, task-aware radar configurations that match or exceed full-array baselines while utilizing fewer receivers, potentially reducing radar costs and integration complexity. These results indicate that the learned optimal MIMO radar design is dependent on the fusion stack and the downstream perception task.
Blogger's Review: DeeperRadar introduces a novel approach by incorporating a learnable MIMO design module, significantly enhancing the efficiency and cost-effectiveness of autonomous perception. This innovation not only advances perception technology but also lays a foundation for future intelligent transportation systems.