RADEN: An Automotive Millimeter-Wave Radar Point Cloud Densifying Network for 3-D Object Detection

The development of Internet of Things (IoT) technology in the field of intelligent vehicle, has raised requirements on the measurement effect of automotive millimeter wave (MMW) radar sensor. However, the extremely sparse characteristic of automotive MMW radar point cloud (RPC) leads to insufficient motion states measurement information of 3D objects, compromising perception reliability in intelligent vehicle systems. In this article, an automotive millimeter wave radar point cloud densifying network for 3D object detection (RADEN) is proposed. Firstly, Multi-Level Vertical Point Cloud Transformer (MLVPT) module is designed through decompose-encode-aggregate (DEA) strategy, which enhances the Transformer’s capacity to learn sparse RPC’s local-global geometrical features. Secondly, Edge-Wise Feature (EWF) module is designed to extract the edge feature avoiding contour distortion based on coordinate-wise of Manhattan distances. Thirdly, Dynamic Scanning Regression (DSR) module is designed to generate dense and high-quality RPC by enhancing the relative geometric relation of neighbor points. The scanning operator in this module is designed to obtain vector information through two considerations, one is inducing learnable vector parameter to fit local geometries, the other is adopting adjusted cosine similarity as weights. Lastly, Feature Kullback-Leibler Divergence Loss item (FKLD-Loss) is proposed to improve the reliability of level-wise feature. This loss item is proposed based on the asymmetry characteristic of KLD to align the features of extraction and the baseline unidirectionally. In the experiment, RADEN is evaluated on the View-of-Delft (VoD) dataset and the TJ4DRadSet dataset with different 3D object detectors. The results show that RADEN generates dense and high-quality RPC for 3D object detection task.