Feature Fusion and Interaction Network for 3D Object Detection based on 4D Millimeter Wave Radars

The 4D Millimeter wave (mmWave) radar has demonstrated excellent application prospects in autonomous driving due to its economic efficiency and robustness in harsh weather conditions. It provides a reliable solution to cope with extreme environments. However, the 4D mmWave radar contains lots of noise and provides sparser point clouds than LiDAR. So significant challenges are posed for 3D object detection based on 4D mmWave radar. In this paper, we propose a new network named Feature Fusion and Interaction Network (FFI-Net) for 4D radar-based 3D object detection. Considering the sparsity of 4D radar point clouds, we introduce a new encoder which taking into account the interaction between features of different dimensions in the raw data. Additionally, we propose a feature interact module to fuse feature map at different scales. Extensive experiments on View-of-Delft (VoD) dataset shows that our FFI-Net can achieve considerably superior performance compared to some state-of-the-art methods.