Abstract: To address the lack of explicit depth constraints in 2D image features and object location priors for 3D query initialization in vision-based 3D object detection for autonomous driving, we propose DG-3D, a depth-guided 3D object detection method that introduces depth information into feature encoding and query initialization. First, a global camera-aware depth prediction module uses camera intrinsics to explicitly regress a pixel-wise depth map. Second, a geometric encoding module constructs planar geometric and depth-distance priors, and uses depth-aware geometric attention to enhance the spatial awareness of visual features. Finally, an object-center query generation module combines depth-guided 2D detection results with predicted depths and generates object-center 3D queries through back-projection, providing object location priors for query initialization. Experiments on the large-scale nuScenes autonomous driving dataset show that DG-3D outperforms existing state-of-the-art methods across all configurations. With ResNet101+ as the backbone, DG-3D achieves 52.8% mean average precision (mAP) and a 61.0% nuScenes detection score (NDS).