A 3D Object Detection Method Based on Cross-Attention Multimodal Feature Fusion

As a key component of the environmental awareness system of autonomous vehicles, efficient and accurate 3D detection of obstacles is directly related to driving safety. However, in complex and ever-changing driving environments, single sensory input like camera or LiDAR often fails to provide accurate and comprehensive environmental information. To improve the effectiveness of 3D object detection, this paper focuses on a multi-perception feature fusion strategy and proposes a supervised learning method called MPF-Net. Firstly, the image features are mapped into a 3-dimensional view cone space, then fused with the LiDAR features in a unified bird's-eye-view (BEV) space, finally an efficient BEV network is designed for 3D object detection. The performance has been validated on KITTI database and compared with several other models, the results demonstrate that the proposed method has more accuracy.