Scholarlink

<p>Three-dimensional (3D) object detection is essential for autonomous driving (AD) and advanced driver assistance systems (ADAS). However, most existing 3D object detection works primarily focus more on the detection accuracy than on the detection efficiency, although the latter is also a critical metric for practical applications. In this paper, we propose RadarNeXt, a lightweight and real-time 3D object detector based on 4D millimeter-wave (mmWave) radar point clouds. We introduce a re-parameterization mechanism in the feature extraction process to enhance inference efficiency while maintaining the multi-branch architecture’s capability to construct multi-scale features. Moreover, to emphasize the irregular foreground features of radar point clouds, and to suppress background clutter, we present a Multi-path Deformable Foreground Enhancement Network (MDFEN), ensuring high detection accuracy while minimizing the trade-off in speed and limiting the number of extensional parameters. Experimental results show that RadarNeXt can achieve the outstanding efficiency performance of over 67 frames per second (FPS) on a RTX A4000 GPU and over 28 FPS on the Jetson AGX Orin. Notably, RadarNeXt meets the real-time inference requirements while achieving more than 50 mean average precision (mAP) on the View-of-Delft (VoD) dataset and over 32 mAP on the TJ4DRadSet (TJ4D) dataset. This research demonstrates that RadarNeXt represents a novel and effective paradigm for 3D object detection based on 4D mmWave radar.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Radarnext: lightweight and real-time 3D object detector based on 4D mmWave imaging radar

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月29日

查看全文
  • Liye Jia,
  • Runwei Guan,
  • Haocheng Zhao,
  • Qiuchi Zhao,
  • Ka Lok Man,
  • Jeremy Smith,
  • Limin Yu,
  • Yutao Yue

摘要

Three-dimensional (3D) object detection is essential for autonomous driving (AD) and advanced driver assistance systems (ADAS). However, most existing 3D object detection works primarily focus more on the detection accuracy than on the detection efficiency, although the latter is also a critical metric for practical applications. In this paper, we propose RadarNeXt, a lightweight and real-time 3D object detector based on 4D millimeter-wave (mmWave) radar point clouds. We introduce a re-parameterization mechanism in the feature extraction process to enhance inference efficiency while maintaining the multi-branch architecture’s capability to construct multi-scale features. Moreover, to emphasize the irregular foreground features of radar point clouds, and to suppress background clutter, we present a Multi-path Deformable Foreground Enhancement Network (MDFEN), ensuring high detection accuracy while minimizing the trade-off in speed and limiting the number of extensional parameters. Experimental results show that RadarNeXt can achieve the outstanding efficiency performance of over 67 frames per second (FPS) on a RTX A4000 GPU and over 28 FPS on the Jetson AGX Orin. Notably, RadarNeXt meets the real-time inference requirements while achieving more than 50 mean average precision (mAP) on the View-of-Delft (VoD) dataset and over 32 mAP on the TJ4DRadSet (TJ4D) dataset. This research demonstrates that RadarNeXt represents a novel and effective paradigm for 3D object detection based on 4D mmWave radar.