Enhancing Dynamic Visual SLAM Precision Based on Pseudo-Semantic Segmentation and Features Weighting

Localization accuracy in dynamic Simultaneous Localization and Mapping (SLAM) is degraded when moving objects violate the static-environment assumption. This paper proposes a dynamic Visual SLAM (VSLAM) method that integrates pseudo-semantic segmentation with weighted optimization to address the issue. First, an adaptive feature extraction module is designed, which establishes a Gaussian distribution model for intra-frame features and combines it with an inter-frame matching quality function to overcome the limitations of traditional feature extraction methods. Second, a depth-based pseudo-semantic segmentation method is proposed, where the K-means clustering algorithm is employed to perform clustering analysis on depth information, generating accurate dynamic object masks for precise exclusion of dynamic features. Finally, a weighted optimization model based on epipolar constraints is constructed. By calculating the distance from feature points to epipolar lines and assigning weights accordingly, the model suppresses interference from dynamic feature points while enhancing the contribution of high-confidence static feature points. Experimental results demonstrate that, compared to ORB-SLAM3, the proposed method reduces the root mean square error (RMSE) of the Absolute Trajectory Error (ATE) by approximately 98.5% on the TUM dataset, while outperforming other existing state-of-the-art methods in terms of localization accuracy and robustness.