A Light Architecture for Radar Range–Doppler Detection

Millimeter-wave (MMW) radar, as a widely used sensor, is increasingly employed in intelligent driving systems. Particularly in challenging driving conditions, such as adverse weather or varying light levels, radar often outperforms other sensors like camera and LiDAR. However, due to the relatively low resolution of radar sensors and its original complex data being difficult to process, deep learning approaches have been less explored in radar target detection compared to camera and LiDAR. Furthermore, we observe that existing methods in the literature often suffer from slow processing speeds and relatively low precision, which poses significant challenges for meeting the detection accuracy and real-time performance requirements of autonomous driving systems. To address these challenges, this paper proposes a lightweight radar target detection network designed to efficiently detect targets from FFT-processed radar frequency data in Range-Doppler spectrum format. The proposed model draws inspiration from YOLOv7-tiny and is further optimized to accommodate the specific characteristics of radar data. A novel, lightweight feature extraction backbone and a multi-scale feature enhancement network are introduced, significantly improving detection accuracy while reducing computational complexity. Additionally, to fully exploit radar data, various data augmentation strategies are explored, and a more efficient augmentation method is developed. Extensive experimental evaluations demonstrate that our model achieves superior detection accuracy and faster inference speed compared to state-of-the-art methods on the RADDet and CARRADA datasets.