Pre-Training For mmWave Radar Object Detection Through Masked Image Modeling

Millimeter-wave (mmWave) radar object detection (ROD) is an indispensable technique in autonomous driving technology because of its ability to maintain stability in complex environments. Deep learning (DL) techniques have made promising achievements in radar target detection techniques, but these techniques usually require large labeled datasets. However, the labeling of millimeter-wave radar datasets is more complex. In this paper, we introduce a novel pre-training method based on mask image modeling (MIM) in the ROD task, aiming to address the dataset limitations in the development of ROD techniques. The method performs spatial and temporal masking operations on unlabeled millimeter-wave radar images, which can acquire more spatial and temporal features simultaneously in the pre-training stage. Furthermore, inspired by T-RODNet, this paper presents a lightweight SS-RODNet network for testing the effectiveness of the proposed MIM method. The model aims to fully integrate the advantages of extracting local and global features. Numerous ablation experiments prove that the proposed MIM method is effective, and SS-RODNet achieves second place on the CRUW datasets.