RadarPillarDet: Multi-Pillar Feature Fusion with 4D Millimeter-Wave Radar for 3D Object Detection

<div class="section abstract"><div class="htmlview paragraph">Object detection is one of the core tasks in autonomous driving perception systems. Most perception algorithms commonly use cameras and LiDAR sensors, but the robustness is insufficient in harsh environments such as heavy rain and fog. Moreover, velocity of objects is crucial for identifying motion states. The next generation of 4D millimeter-wave radar retains traditional radar advantages in robustness and speed measurement, while also providing height information, higher resolution and density. 4D radar has great potential in the field of 3D object detection. However, existing methods overlook the need for specific feature extraction modules for 4D millimeter-wave radar, which can lead to potential information loss. In this study, we propose RadarPillarDet, a novel approach for extracting features from 4D radar to achieve high-quality object detection. Specifically, our method introduces a dual-stream encoder (DSE) module, which combines traditional multilayer perceptron and attention-based methods. The DSE module serves as a powerful point feature extractor that enhances feature dimensions. Compared to other methods, Sum-Avg-Max Pillar Encoding (SAMPE) module effectively enriches the features of sparse radar point clouds by collecting various pillar features using three different encoders. Additionally, to effectively address the issue of noise points in 4D radar, the designed multi-pillar self-attention (MPSA) module can adaptively learn the weights of different pillar features, thereby enhancing the quality of the 4D radar bird's eye view (BEV) features. Experimental results on the View of Delft (VoD) dataset show that the proposed RadarPillarDet achieves excellent detection performance, with a performance 3.22% mAP higher than the baseline.</div></div>