<p>Dense visual SLAM systems based on 3D Gaussian splatting have successfully achieved photorealistic scene reconstruction. However, their performance is limited in efficiency and global consistency of camera tracking and mapping. To address these issues, we propose KBGS-SLAM, a keyframe-optimized and bundle-adjusted dense visual SLAM system. Specifically, we design a robust keyframe extraction and management method to select valuable keyframes and maintain a keyframe window throughout the SLAM process. Subsequently, to accelerate the SLAM processing, we perform Gaussian densification and optimization only on keyframes and use a binary mask when calculating the loss. Furthemore, to reduce the cumulative errors and maintain multi-view consistency, we efficiently incorporate local bundle adjustment and global bundle adjustment to optimize 3D Gaussians as well as camera poses. Finally, compared to state-of-the-art visual SLAM methods, KBGS-SLAM reduces the camera tracking error by 7% and 23% on the TUM RGB-D and Replica datasets, respectively. It also improves PSNR by 5% on the Replica dataset, achieving a high frame rate.</p>