To improve the accuracy and robustness of 3D object detection from a Bird's-Eye View (BEV) in multi-camera visual perception for autonomous driving, this paper proposes a multi-view image object detection method based on an improved BEVFormer. To introduce perspective-based regional auxiliary supervision at the image feature level, using BEV annotations to impose projection constraints on the image view, thereby guiding the backbone network to learn fine-grained local features for 3D spatial localization. This paper establishes instance-level consistency constraints, at the BEV representation level and explicitly enforces the consistency of BEV features for the same physical object across different viewpoints based on a contrastive learning mechanism, thereby mitigating instance-level feature drift caused by view transformations. Furthermore, this paper trains and evaluates the proposed method using the nuScenes dataset and conduct a comparative analysis with various typical multi-view 3D object detection algorithms. The experimental results show that, compared to the BEVFormer baseline model, the proposed method achieves improvements of 1.6 percentage points and 0.4 percentage points in mean Average Precision (mAP) and the Normalized Detection Score (NDS), respectively, while reducing the average localization error by 8.7 percentage points, thereby validating the effectiveness of this method for the BEV detection task using multi-view images.