Enhancing Visual-Inertial SLAM via Dynamic-Aware Soft Suppression in DBA-Fusion

Abstract

Visual-Inertial SLAM (VI-SLAM) systems are sensitive to dynamic objects because moving foreground regions may corrupt feature tracking, optical-flow estimation, and subsequent pose optimization. To mitigate this problem, this paper presents a dynamic-aware image-domain soft suppression frontend for DBA-Fusion. The proposed method uses YOLOv8seg to generate pixel-level pedestrian masks and computes a frame-wise dynamic ratio from the masked pixels. This ratio is mapped to an adaptive attenuation coefficient, which is applied to the masked image regions before feature extraction and tracking. Unlike hard masking, the proposed soft suppression reduces the visual influence of dynamic pedestrians while retaining weak texture cues that may help maintain optical-flow continuity.The proposed module is implemented as a lightweight frontend preprocessing strategy and does not modify the internal dense bundle adjustment objective of DBA-Fusion. Experiments on the TUM-VI corridor1 sequence show that the proposed adaptive soft suppression improves trajectory completeness and path-length consistency while maintaining comparable absolute trajectory accuracy. Supplementary mask analysis on TUM RGB-D walking sequences further verifies that pixel-level pedestrian masks can provide useful dynamic-region priors for frontend image modulation. Ablation studies analyze the influence of hard masking, fixed soft suppression, adaptive attenuation ranges, and runtime overhead under offline mask conditions.