MARDF:Mining and Reconstructing 4D-Radar Derived Features in Point Cloud for Object Detection

Detecting objects plays an essential role in autonomous driving systems. Recently, 4D millimeter-wave radar, which captures spatial information in (x, y, z, v) coordinates along with velocity v, has proven to produce denser point clouds compared to traditional 2+1D radar systems that only provide position and velocity information. While most current point cloud detection techniques are tailored to LiDAR data, they may not be directly applicable to radar data, which is typically sparse, noisy, and includes velocity components. In this work, we propose a novel method to enhance radar-derived point cloud features for object detection, referred to as MARDF. This approach leverages the radar point cloud's density approximation and incorporates point velocities, which boosts the dimensionality of the point cloud features. Experiments conducted using the TJ4DRadSet and VoDataset in various traffic environments indicate that the proposed object detection method outperforms classical point cloud detection methods. Furthermore, this module is adaptable to other detection methods and demonstrates an improvement in accuracy.