IPD-SLAM: A VSLAM Algorithm Based on Instance Segmentation and Parallax Angle Modeling in Dynamic Scenes

Abstract

In order to solve the problems of existing SLAM algorithms such as incomplete segmentation of dynamic objects and difficulty in accurately judging the motion state of potential dynamic objects, IPD-SLAM is proposed. To minimize the impact of dynamic objects on the localization accuracy of the SLAM system, the algorithm precisely labels potential dynamic objects by employing a residual coordinate attention module. This module is designed to direct the model’s focus toward regions of the image where segmentation has been overlooked and simultaneously integrates an external relative position bias to refine the boundary semantics of these missed segmentation objects. Additionally, we introduce a dynamic object judgment strategy based on the parallax angle model to predict the motion states of potential dynamic objects more effectively. Evaluation on publicly available datasets demonstrates that IPD-SLAM significantly outperforms existing methods, achieving average ATE RMSE reductions of 48.48%, 26.64%, and 17.72% compared to DS-SLAM, Crowd-SLAM, and DynaSLAM algorithms, respectively.

Keywords

• Simultaneous localization and mapping (SLAM)
• dynamic scenarios
• instance segmentation
• motion judgment