3D image reconstruction using an improved BEV model and global convolutional attention fusion

In autonomous driving and computer vision, 3D object detection plays a critical role but faces challenges related to the effective extraction and integration of multi-view features. The existing BEVFormer model, which uses CNNs to convert images into a bird's-eye view (BEV), shows potential but struggles to capture fine-grained details and multi-scale information, especially in high-resolution, complex scenes. To address these limitations, we propose the MultiCAN-DEBEV model, which integrates the MSF-DySample, GCAF, and MSDE modules. These modules improve the handling of multi-scale features, enhance feature expressiveness, and strengthen detail representation. Experiments on the nuScenes dataset show significant performance improvements, and the modular design ensures broad adaptability to other 3D detection models.