Vision-IMU multi-sensor fusion semantic topological map based on RatSLAM

Abstract

The simultaneous localization and mapping (SLAM) method based on the brain-space cognitive model can improve the localization accuracy of the robot through memory. The traditional method has poor robustness in complex environments such as intense motion, weak texture, and illumination changes, and lack of understanding of the environment. Our work is based on RatSLAM. To improve the robustness, we propose a multi-sensor fusion algorithm, which integrates a camera with an IMU. To improve the understanding of the environment, we use the Yolo to extract the semantic information of objects and store it in the topological nodes and construct a 2D topology map. Yolo-v4 is used to replace Yolo-v3 to improve the accuracy of semantic tag extraction. We validated our approach in the pubic dataset and real environments. The mean RMSE APE is reduced by 51.716% compared to RatSLAM. Also, our method can build a semantic topological map of the environment.