EFA-RadNet: Efficient Feature Aggregation with Balanced Attention for Raw Radar Multi-Task Learning

Original high-definition radar data contains rich environmental information, including distance, Doppler velocity, and azimuth. However, extracting robust features from such sparse and noisy frequency-domain data remains a challenge. To address this issue, this paper proposes an improved multi-task network, the Efficient Feature Aggregation with Balanced Attention Radar Network (EFA-RadNet). This network introduces the VoVNetV2 architecture into the field of raw radar perception and effectively preserves feature diversity across different receptive fields through a One-Shot Aggregation (OSA) module, avoiding signal aliasing. In addition, we propose an attention mechanism module, Balanced effective Squeeze–Excitation (B-eSE), which is better suited for sparse radar processing and effectively addresses the problem of weak target loss in the radar spectrum. Experiments on the RADIal dataset show that our EFA-RadNet achieves excellent target detection performance while also attaining optimal accuracy in free space segmentation.