<p>Modern depth sensors face limitations including low resolution, narrow FOV, and dynamic interference, leading to missing data and distortions. Existing methods mainly process single-frame depth features, underutilizing temporal cues like motion trajectories and occlusions in sequential frames. To address these limitations, we propose a three-stage Guidance-Calibration-Optimization architecture. Spatio-temporal guidance network employs memory units to process sequential frames and generate initial guided depth maps; depth-dominant network performs multi-modal calibration on the guided depth map, RGB semantics, and raw depth map to suppress modal interference and reinforce reliable features; geometric optimization smooths noise in homogeneous regions while preserving boundaries, improving visual consistency. This design reduces optical flow dependency and facilitates depth-focused temporal modeling. Experiments indicate enhanced depth video quality with improved completeness and accuracy, supporting dynamic scene understanding.</p>