Current mainstream monocular 3D object detection networks,which are based on keypoint detection,exhibit limitations in temporal feature modeling,geometric constraint utilization,and depth estimation,thereby constraining the overall performance of 3D detectors.This paper proposes MonoTGD (Monocular Temporal Geometric Deep),an improved monocular 3D object detection algorithm.The proposed framework incorporates three key modules:a temporal feature interaction module that leverages both long- and short-term temporal information from multi-frame sequences to enhance feature representation consistency and dynamic modeling capabilities;a geometric structure enhancement module that improves keypoint prediction accuracy by expanding the keypoint set and enforcing geometric consistency constraints;a pseudo-depth generation and supervision module that produces pseudo-depth maps without requiring LiDAR data,thereby providing effective supervisory signals for depth estimation.Crucially,all these modules are used only during the training phase,introducing no additional computational cost during inference.Experiments on the KITTI3D dataset show that MonoTGD significantly improves performance.Specifically,it increases the average precision of 3D detection on the validation set by 4.25 percentage points in the easy category.More importantly,it achieves a gain of 4.82 percentage points in the moderate difficulty category on the test set,which fully validates the method's effectiveness in practical application scenarios.