Dense Semantic Mapping Method for Dynamic Workshop Environments

The intelligentization of workshops is a key direction for the smart transformation of the manufacturing industry. Autonomous navigation of intelligent mobile robots in dynamic workshop environments raises higher requirements for environmental perception and map building. Therefore, this paper proposes a dense semantic map construction algorithm suitable for dynamic workshop scenarios. The algorithm is based on the ORB- SLAM2 framework, combined with a lightweight semantic segmentation network, dynamic target elimination technology, and a dense semantic mapping thread, significantly improving the system's localization accuracy, map-building precision, and robustness. A dynamic elimination strategy is proposed by integrating optical flow, multi-view geometry constraints, and region growing methods with semantic priors, effectively reducing the interference from dynamic objects. Experimental results show that, compared to ORB-SLAM2, the algorithm proposed in this paper reduces the RMS error (RSME) and standard deviation (Std) of the Absolute Trajectory Error (ATE) by 82.02% and 81.26%, respectively, and the RMS error and standard deviation of Relative Pose Error (RPE) by 39.40% and 36.23%, respectively, on the TUM public dataset. By utilizing keyframe dynamic region elimination and generating point clouds based on poses, semantic information is embedded into the dense map. Furthermore, the map's readability and clarity are optimized through the use of an octree structure, resulting in an accurate, robust, and semantically enriched dense map, which provides strong support for workshop intelligentization and the execution of high-level robotic tasks.