Jia Liye ; Guan Runwei ; Zhao Haocheng ; Zhao Qiuchi ; Man Ka Lok ; Smith Jeremy ; Yu Limin ; Yue Yutao

Abstract

3D object detection is crucial for Autonomous Driving (AD) and AdvancedDriver Assistance Systems (ADAS). However, most 3D detectors prioritizedetection accuracy, often overlooking network inference speed in practicalapplications. In this paper, we propose RadarNeXt, a real-time and reliable 3Dobject detector based on the 4D mmWave radar point clouds. It leverages there-parameterizable neural networks to catch multi-scale features, reduce memorycost and accelerate the inference. Moreover, to highlight the irregularforeground features of radar point clouds and suppress background clutter, wepropose a Multi-path Deformable Foreground Enhancement Network (MDFEN),ensuring detection accuracy while minimizing the sacrifice of speed andexcessive number of parameters. Experimental results on View-of-Delft andTJ4DRadSet datasets validate the exceptional performance and efficiency ofRadarNeXt, achieving 50.48 and 32.30 mAPs with the variant using our proposedMDFEN. Notably, our RadarNeXt variants achieve inference speeds of over 67.10FPS on the RTX A4000 GPU and 28.40 FPS on the Jetson AGX Orin. This researchdemonstrates that RadarNeXt brings a novel and effective paradigm for 3Dperception based on 4D mmWave radar.