MAFF-Net: Enhancing 3D Object Detection With 4D Radar via Multi-Assist Feature Fusion

Abstract

Perception systems are crucial for the safe operation of autonomous vehicles, particularly for 3D object detection. While LiDAR-based methods are limited by adverse weather conditions, 4D radars offer promising all-weather capabilities. However, 4D radars introduce challenges such as extreme sparsity, noise, and limited geometric information in point clouds. To address these issues, we propose MAFF-Net, a novel multi-assist feature fusion network specifically designed for 3D object detection using a single 4D radar. We introduce a sparsity pillar attention (SPA) module to mitigate the effects of sparsity while ensuring a sufficient receptive field. Additionally, we design the cluster query cross-attention (CQCA) module, which uses velocity-based clustered features as queries in the cross-attention fusion process. This helps the network enrich feature representations of potential objects while reducing measurement errors caused by angular resolution and multipath effects. Furthermore, we develop a cylindrical denoising assist (CDA) module to reduce noise interference, improving the accuracy of 3D bounding box predictions. Experiments on the VoD and TJ4DRadSet datasets demonstrate that MAFF-Net achieves state-of-the-art performance, outperforming 16-layer LiDAR systems and operating at over 17.9 FPS, making it suitable for real-time detection in autonomous vehicles.