MSVDNet: A multi-scale vehicle detection network for target detection in autonomous driving

With the development of new energy vehicle technology, the demand for target detection in autonomous driving scenarios has grown. Synthetic aperture radar image technology combined with deep learning can replace traditional remote sensing target recognition. However, detecting objects in SAR images for autonomous driving faces challenges like small vehicle targets and varying scales. To address these, this paper proposes MSVDNet, a method based on lightweight YOLOv5 for better multi-scale object detection in SAR images. It constructs two key modules: a cross-stage multi-scale receptive field feature extraction module with enhanced feature representation capability, and a feature adaptive fusion pyramid module with learnable fusion coefficients. Compared with existing methods, MSVDNet shows significant improvements. Experimental results on SSDD and Berkeley DeepDrive datasets demonstrate its superiority: it achieves 61.1 % AP, which is higher than OTA’s 59.1 % and outperforms YOLOv5s. With 24.5 GFLOPs, it reduces computational load by 29 % compared to the Res2Net baseline. Notably, it enhances small-target detection with 55.4 % APS, which is 3.3 % higher than YOLOv5s, while enabling real-time inference at 24.2 ms on embedded hardware.