Existing 3D Gaussian Splatting methods suffer from limited accuracy in camera pose estimation under complex scenes and rely on insufficient depth supervision. To address these issues, this paper proposes an improved framework that jointly optimizes camera poses and enforces depth consistency constraints, enhancing the robustness and fidelity of 3D reconstruction. To enable end-to-end camera pose correction, we perform simultaneous optimization of camera position and rotation parameters via differentiable optimization, guided by a regularization term based on the initial poses. In addition, we propose an edge-aware depth regularization strategy that leverages image gradients to compute edge weights for refining depth distortion maps. To further improve depth accuracy, we introduce a multi-scale depth consistency constraint mechanism. Experimental results on the Mip-NeRF 360, Tanks and Temples, and Deep Blending datasets showcase the superior performance of the proposed method.