VSLAM algorithm based on instance segmentation and motion consistency constraints in dynamic scenes

The current visual simultaneous localization and mapping (SLAM) algorithm is prone to cause problems such as difficulty in completely marking potential dynamic objects and inability to accurately judge the motion state of dynamic objects in dynamic scenes, which affects the accuracy of camera pose estimation. Therefore, a visual SLAM (VSLAM) algorithm based on instance segmentation and motion consistency constraints in dynamic scenes is proposed. A ReT-encoder module is designed to focus on and extract the local and global feature information of the image. At the same time, a mixed weight of feature pyramid network (FPN) module is designed to accurately mark potential dynamic object regions by fusing and learning large object features and small object features through weight coefficients. In order to furtherly reduce the impact of dynamic objects on the positioning accuracy of SLAM system, the motion state of the object is judged and the dynamic objects are eliminated through pose estimation and deflection error. The results of validation on the public dataset TUM and real scenarios show that the average absolute trajectory root mean square error of the proposed algorithm is reduced by 95.23%, 46.85%, and 15.88% compared to ORB-SLAM2, DS-SLAM, and DynaSLAM algorithms, respectively. © 2023 Editorial Department of Journal of Chinese Inertial Technology. All rights reserved.