DynaFusion-VIO: A Hybrid Visual Inertial Approach for Dynamic World Navigation

This paper presents an advanced Visual–Inertial Odometry SLAM (VIO-SLAM) system designed to deliver robust and accurate performance in dynamic environments. The proposed front-end employs a deep learning–based visual odometry model capable of identifying and handling dynamic objects, thereby preventing their adverse influence on the motion estimation process. To further enhance robustness, the system integrates this visual front-end with a reliable inertial fusion module inspired by PVGO, enabling consistent trajectory estimation even in challenging conditions. The core contribution of this research lies in the intelligent integration of a deep learning–driven dynamic-aware front-end with an optimization-based inertial back-end. This combination results in a stable and precise estimation framework suitable for complex and crowded environments. Experimental results demonstrate that the proposed system significantly outperforms existing approaches in terms of Average Trajectory Error (ATE) and overall system stability, particularly in scenarios where traditional methods struggle due to dynamic elements. The simulation codes and implementation details will be made publicly available at the GitHub repository below upon the publication of the paper: https://github.com/rezamj201/DynaFusion-VIO.