CAN-FFTRadNet: Lightweight Attention and Weighted Fusion for Multi-task Detection and Free-Space Segmentation on DDMA-MIMO Radar

To address the issues of insufficient noise suppression and false detections in the FFTRadNet model for target detection in DDMA-MIMO millimeter-wave radar and free-space segmentation in autonomous driving, this paper proposes an improved model, CAN-FFTRadNet, which integrates a cascaded attention mechanism with the lightweight MobileNetV3 network. This model replaces the original feature extraction network with MobileNetV3, effectively reducing the number of parameters and computational complexity. It combines the ECA + SA cascaded attention mechanism to enhance both channel and spatial feature representations. Additionally, the Weighted Feature Fusion with Sparse Coding (WFF-SC) method is introduced for multi-scale feature fusion to improve integration efficiency. Experimental results show that in the tasks of object detection and free-space segmentation, the CAN-FFTRadNet method achieves improved recall rate and segmentation quality while maintaining comparable precision. Specifically, its Average Recall (AR) and mean Intersection over Union (mIoU) are increased by 7.1% and 5.9%, respectively. Furthermore, the reduced model parameters make it more suitable for deployment on embedded platforms.