A Static Feature Point Extraction Algorithm for Visual-Inertial SLAM

Aiming at the fact that most simultaneous localization and mapping (SLAM) methods are fragile in dynamic environment, a static feature point extraction algorithm for visual-inertial SLAM is proposed. Firstly, the distance from feature matching to transformation matrix is calculated by cleverly combining inertial measurement unit (IMU) data and visual information. Secondly, a static probability model for distinguishing feature point categories is established based on the binomial logistic regression model. These categorical feature points are used in subsequent algorithms to select more stable data associations. Finally, the algorithm is integrated into the ORBSLAM-3 system, and extensive experiments are performed on the OpenLORIS-Scene public dataset and real world to verify the performance of the integrated system. The results show our algorithm not only maintains the high-precision localization performance of the original system in the static environment, but also improves the localization accuracy of the original system in the dynamic environment.