Visual SLAM algorithm based on semantic information and IMU fusion in dynamic scenes

Visual SLAM algorithms usually assume that objects in the environment are static or low-motion, and perform poorly in dynamic scenes due to the influence of dynamic objects. To address this issue, this paper proposes a SLAM algorithm based on semantic information and IMU fusion. This algorithm is an improvement based on the ORB-SLAM2 framework. Firstly, we use the real-time instance segmentation model YOLACT to identify potential dynamic objects in the image frame, and then further detects and identifies dynamic objects to eliminate dynamic feature points. Secondly, an extended Kalman filter motion model is established. The camera pose is calculated using IMU data that is not affected by the environment, and then fused with the pose calculated based on visual information to further reduce the adverse effects of the dynamic environment. Experiments show that when this algorithm is running in a dynamic scene, compared to pure visual SLAM algorithms, our algorithm has greatly improved accuracy and robustness.