E-RODNet: Lightweight Approach to Object Detection by Vehicular Millimeter-Wave Radar

The main sensors employed for achieving autonomous driving perception tasks include cameras, LiDAR, and millimeter-wave radar. Millimeter-wave radar offers the advantages of all-weather and all-day operation, helping to compensate for the shortcomings of cameras and LiDAR in adverse environmental conditions and ensuring the safety of autonomous driving technology. Currently, most object perception models based on millimeter-wave radar are overly complex, requiring substantial computational and storage resources during inference, which cannot meet the real-time requirements of autonomous driving. In light of this, this article proposes a more efficient new model, referred to as E-RODNet. The model employs the ConvFormer block as the primary feature extraction unit and adjusts the SepConv operator for spatiotemporal data modeling. To enhance the model’s global modeling capability, a global feature fusion module is introduced, adding global context information to each pixel after encoding features. In addition, the short-term sequence feature fusion module is improved based on the characteristics of the CRUW dataset. The application of these two techniques allows the model to achieve state-of-the-art performance with minimal parameters and computational resources. Compared to T-RODNet, E-RODNet achieves competitive performance with only 3.8% of the model parameters, improving average precision (AP) by 1.63%, and requiring only 18.2% of the GFLOPs of T-RODNet. Our model and training code can be available at https://github.com/lupeng-xm/E-RODNet.