Three-dimensional dynamic uncertainty semantic SLAM method for a production workshop

Abstract

Simultaneous localization and mapping (SLAM) is the basis for intelligent navigation and intelligent security of mobile robots in a workshop. However, there are numerous dynamic objects in a workshop, such as robots and operators, which can decrease the accuracy of robot localization and mapping. To solve this problem, this paper proposes a three-dimensional dynamic semantic system named the PW_SLAM. First, the production workshop-oriented lightweight semantic segmentation network (PWnet) is integrated with the SLAM system by a robot operating system(ROS) to provide semantic information. Next, a dynamic uncertainty keypoint classifier(DUKC) is proposed to filter the dynamic keypoints and improve the localization accuracy. Then, the map is stored in the form of an octree, and a dynamic object map filter are designed to filter the dynamic objects in the octree map. The proposed system is evaluated on the TUM RGB-D public datasets and the results show that the RMSE of absolute trajectory error in PW_SLAM is decreased by 92.80%, 98.85%, 87.21% compared to that of ORB_SLAM3 on the TUM highly dynamic datasets “fr3/walking_xyz”, “fr3/walking_static”, and “fr3/walking_rpy”, respectively. In addition, the PW_SLAM is also applied to the real workshop environment, and the results show that it can effectively eliminate dynamic objects and build a static octree map of a workshop environment effectively.