Gaussian Simultaneous Localization and Mapping (SLAM) technology has problems with reconstruction accuracy and visual quality in complex scenes and high-frequency details. Therefore, in this paper, a dense visual SLAM framework based on Gaussian Splatting is proposed to represent the scene as 3D Gaussian primitives. A new adaptive sampling method can be used to select multiple representative samples from the data for reconstruction. It is an exponential decay weight method that gives higher weight to recent observations to improve reconstruction accuracy and tracking performance. Add a perception loss function to the system that emphasizes high-level semantic features in optimisation. Strengthen the reconstruction of fine scene details and improve the general visual quality this way. According to the above experiments, the proposed method can achieve good results in scene reconstruction and precise camera trajectory estimation. Based on the above comparison, PSNR is approximately 32.38% higher on average, SSIM has risen by 17.48% on average, LPIPS has increased by 47.57% on average, and Absolute Trajectory Error (ATE) (Root Mean Square Error, RMSE) has improved by about 29.20% on average.
