: The conventional assumption of environmental staticity constrains the performance of established visual SLAM frameworks, such as ORB-SLAM3, resulting in notable localization drift within dynamic scenarios. To address this limitation, this study proposes a dynamic visual SLAM architecture integrating temporal-semantic information fusion and adaptive soft-weight optimization. In the proposed approach, feature-level variations are modeled as probabilistic random variables. Initial semantic information is extracted using YOLOv11, and Bayesian joint inference is employed to fuse these semantic priors with geometric consistency across multiple frames. Dynamic probabilities are quantitatively estimated by comparing reprojection residuals against a Chi-square statistical distribution, incorporating temporal recursive filtering to enhance robustness. Rather than utilizing binary hard thresholds, the method adopts continuous dynamic confidence scores within the information matrix of the backend pose-graph optimizer, thereby adaptively suppressing dynamic outliers while preserving static geometric constraints. Comprehensive ablation studies are conducted to evaluate the individual contributions of each algorithmic module. Experimental results on the TUM RGB-D walking_xyz sequence demonstrate that the proposed system reduces the Absolute Trajectory Error (ATE) root mean square error