RD Mamba: A Lightweight Radar Range-Doppler Spectrum Object Detection Model

This paper proposes RD Mamba, a lightweight radar Range-Doppler (R-D) spectrum object detection model, designed to enhance the radar system's target detection capability in multi-target and noisy environments. We introduce the Dynamic Spatio-Temporal (DST) Module, which integrates the Dual-Axis Spatial Fusion (DF) Block, SS2D, and Triplet Attention (TA) Block to enhance local feature extraction and global context integration. The DF Block employs a dual-branch convolutional structure to capture robust local features, while the TA Block utilizes a triplet attention mechanism to efficiently aggregate multi-dimensional features. Experiments on the RADDet and CARRADA datasets demonstrate that RD Mamba achieves Mean Average Precision(mAP)50 scores of 68.82% and 62.35%, outperforming state-of-the-art models. With only 4.2M parameters and 9.8G Floating Point Operations Per Second(FLOPs), the model strikes a balance between accuracy and efficiency, making it suitable for real-time applications in resource-constrained environments, such as autonomous driving.