Enhancing Visual-Inertial SLAM with ORB-SLAM3

Simultaneous Localization and Mapping (SLAM) systems face significant challenges in complex environments where moving objects and aggressive motion disrupt accurate mapping and localization. This paper presents an enhanced version of ORB-SLAM3 in monocular IMU mode that addresses these issues through three key modifications. First, the FAST feature detector is replaced with the AGAST feature detector, augmented with adaptive thresholding and adaptive non-maximum suppression (ANMS). This approach dynamically adjusts detection thresholds based on local image statistics and selects keypoints in each image cell by prioritizing high-response features with optimal spatial separation, reducing clustering and ensuring robust, uniform feature distribution across diverse lighting conditions. Second, the default Inertial Measurement Unit (IMU) initialization is replaced with an analytical solution, utilizing a maximum likelihood estimation framework to efficiently and precisely estimate accelerometer bias, gravity direction, and scale factors, thereby improving sensor fusion for visual-inertial SLAM. Finally, the state-of-the-art YOLOv11 object detection model is incorporated to detect and segment dynamic objects, such as pedestrians and vehicles, in real-time, excluding their influence from mapping and localization to enhance system stability in dynamic scenarios. Experimental evaluations on the EuRoC MAV and TUM VI datasets demonstrate substantial improvements in both accuracy and robustness compared to the original ORB-SLAM3 framework.