Abstract:In autonomous driving, the task of using bird’s eye view (BEV) for 3D object detection has attracted significant attention. Existing camera-to-BEV transformation methods are facing challenges of insufficient real-time performance and high deployment complexity. To address these issues, this study proposes a simple and efficient view transformation method that can be deployed without any special engineering operations. First, to address the redundancy in complete image features, a width feature extractor is introduced and supplemented by a monocular 3D detection task to refine the key features of the image. In this way, the minimal information loss in the process can be ensured. Second, a feature-guided polar coordinate positional encoding method is proposed to enhance the mapping relationship between the camera view and the BEV representation, as well as the spatial understanding of the model. Lastly, the study has achieved the interaction between learnable BEV embeddings and width image features through a single-layer cross-attention mechanism, thus generating high-quality BEV features. Experimental results show that, compared to lift, splat, shoot (LSS), on the nuScenes validation set, this network structure improves mAP from 29.5% to 32.0%, an increase of 8.5%, and NDS from 37.1% to 38.0%, an increase of 2.4%. This demonstrates the effectiveness of the model in 3D object detection tasks in autonomous driving scenarios. Additionally, compared to LSS, it reduces latency by 41.12%.