Achieving accurate and efficient 3D object detection from surround-view cameras remains a key challenge in autonomous driving, especially in complex scenarios with constrained computational resources. This paper revisits BEVFormer and introduces two key enhancements to improve accuracy and inference speed. Firstly, an adaptive BEV grid density mechanism is proposed in which deformable convolution dynamically reallocates spatial resolution. Secondly, the BEV encoder is redesigned with depth-wise separable convolutions to replace the standard Self-Attention and FFN stacks. The experiments on datasets nuScenes, Waymo, and KITTI show that the improved BEVFormer achieves 63.2% mAP and 75.1% NDS on nuScenes, outperforming BEVFormer, CenterPoint, and PV-RCNN, while running at 30 FPS on an RTX 3090. Consistent performance improvements across diverse hardware platforms validate the method’s suitability for resource-constrained deployment.