SMURF: Spatial Multi-Representation Fusion for 3D Object Detection with 4D Imaging Radar

• Published in: IEEE Intelligent Vehicles Symposium p. 3141
• Main Authors: Liu, Jianan, Zhao, Qiuchi, Xiong, Weiyi, Huang, Tao, Han, Qing-Long, Zhu, Bing
• Format: Conference Proceeding
• Language: English
• Published: IEEE 02.06.2024
• Subjects: Imaging; Laser radar; Point cloud compression; Radar; Radar detection; Radar measurements; Three-dimensional displays
• DOI: 10.1109/IV55156.2024.10588505

Conventional automotive radar has been extensively utilized in advanced driver assistance systems and autonomous driving, with potential applications in future cooperative perception systems. However, compared to LiDAR-based perception, conventional radar-based perception technologies often encounter limitations such as the absence of elevation information and low resolution. These limitations impede their ability to detect and localize objects in the surrounding environment accurately. In recent years, the development of 4D imaging radar emerged as a promising solution to overcome these limitations. 4D imaging radar can measure the pitch angle, enhancing the understanding of the environment and improving object detection and localization accuracy. Due to the cost-effectiveness and operability in adverse weather conditions of 4D radar, its emergence has attracted attention from both the academic and industrial communities. However, the measurements obtained from 4D radar are subject to noise, primarily stemming from the multi-path propagation of radar signals. Additionally, 4D radar captures less geometry and semantic information than the more dense LiDAR point cloud. As a result, existing 3D object detection algorithms specifically developed for dense LiDAR point cloud may yield suboptimal performance when directly applied to sparse 4D radar point cloud data.