ARC-BEV: Attentive Radar-Camera Fusion 3D Object Detection in Bird-Eye-View Space for Autonomous Driving

In autonomous driving perception system, 3D object detection task plays a crucial role. Recently, although many detection methods with camera only input have shown good performance by retrieving high resolution and rich semantic from images, they are still limited by the lack of depth information. On the other hand, automotive radar, as a common on-board sensor, is typically only employed to perform low-level perception tasks despite its ability to provide accurate depth and doppler velocity information. Therefore, effectively fusing camera and radar for detection can leverage the advantages of both sensors without incurring additional costs. In this paper, we propose a novel Attentive Radar-Camera fusion in Bird-Eye-View (ARC-BEV) 3d detection framework for autonomous driving. Unlike current radar-camera fusion methods, ARC-BEV transforms images into BEV features and fuses them with radar data, thereby mitigating issues of alignment loss and insufficient height information. A spatial attention fusion module is also incorporated to enhance the relevance of features in the spatial domain. ARC-BEV achieves state-of-the-art performance with 43.7% mAP on the nuScenes test set, surpassing our camera-only baseline and existing camera-radar detection methods. Serving as a straightforward and effective framework for the camera-radar fusion detection task, our method demonstrates superior performance and high extensibility and provides benefits for future radar research.