Learning Axis-Aligned Attentions for Efficient and Robust Traffic 3D Object Detection

The Bird’s-Eye View (BEV) framework has emerged as a promising solution for multimodal three-dimensional (3D) object detection in autonomous driving systems. However, the existing methods commonly extract features under the isotropic assumption, ignoring the inherent direction-dependent errors in sensor data. This mismatch results in degraded detection performance in traffic scenarios. To address this issue, we introduce a simple yet effective orientation-sensitive attention module for BEV multimodality fusion networks, termed the Axial Attention Module (AAM). Specifically, we first analyze the characteristics of sensor error to identify the principal axes. Then, we establish interdimensional dependencies through rotation operations and residual transformations, enabling efficient joint encoding of channel and spatial information with minimal computational cost. Finally, we enhance the BEVFusion architecture with our method and name it the BEVFusion-AAM network, which improves the detection accuracy by [Formula: see text] AP[Formula: see text]R40 on the nuScenes dataset with minimal computational overhead. Moreover, we offer comprehensive insights into the performance of the axial attention by visually analyzing GradCAM[Formula: see text] results. Given its lightweight and versatile nature, AAM can be seamlessly integrated into any BEV architecture with insignificant overhead, promising vast engineering application prospects.