<p>Multi-camera based 3D object detection is an important research direction in the field of autonomous driving environment perception. This paper addresses the issues of low detection accuracy and missed detections caused by insufficient image feature extraction and excessive coupling of multi-scale features in the Bird’s-eye view architecture. We propose a novel multi-layer decoupled fusion network model architecture (MLDF-BEV). By designing a multi-layer decoupled fusion network (MLDFNs), we aim to fully extract and fuse multi-layer features to improve object capture capability. Specifically, we introduce a deformable attention mechanism in the backbone network (MLDFNs-Backbone) to enhance the perception and modeling ability of complex spatial structures. Additionally, we design a multi-layer decoupling module (MLDF-module) to decouple the highly coupled multi-scale features. The MLDF-module adaptively adjusts the receptive field through multi-branch information flow, reducing global information blur and local information loss. The fusion and refinement of multi-layer features are achieved through a minimal selection decoupling function, which suppresses background noise in the image. Extensive experiments on the nuScenes and Waymo datasets demonstrate that the MLDF-BEV model achieves superior performance, improving mAP and NDS metrics by 1.8% and 1.5%, respectively, compared to state-of-the-art models. The proposed method has potential applications in the field of autonomous driving environmental perception.</p>