Multiple Attention Mechanism for Camera-Radar Fusion Object Detection

The attention mechanism has gained significant prominence in visual-based object detection due to its outstanding efficiency and adaptability. This paper addresses the challenges of current Attention Mechanisms (AM) in traffic object detection tasks utilizing Camera-Radar Fusion (CRF) technology. A novel Multiple Attention Mechanism for Camera-Radar Fusion Object Detection (MAM-CRF) is proposed. A new network architecture is designed specifically for CRF networks to seamlessly integrate with existing vision-based AM modules. Additionally, an innovative Azimuth Attention Mechanism (AAM) module is developed, combining visual and radar data to enhance the network's focus on azimuthal features. A data scrambling technique is also introduced to reduce over-reliance on individual sensors in fusion-based methods. In the absence of a publicly available dataset tailored to this task, the Camera-Radar University of Washington (CRUW) dataset was selected and modified. Furthermore, a new dataset, the Camera-Radar Nanjing University of Science and Technology Version 1.0 (CRNJUST-v1.0), was created to validate the proposed methodology. Results demonstrate that MAM-CRF improves performance by over 2% with only a marginal complexity increase of 5‰. The method also exhibits robust performance under severe data disturbances. Lightweight and portable, MAM-CRF is well-suited for a wide range of CRF applications.