In context
By early 2023, visual SLAM had become a cornerstone of mobile robotics, yet most mature algorithms assumed static scenes. Dynamic objects—people, chairs, tables—could cause pose estimation errors to spike or even cause system failure, limiting deployment in real-world manufacturing and service environments. This paper, published in the Chinese journal Robot, tackled that gap with a semantic-aware approach.
What was reported
Researchers from Shanghai University proposed an RGB-D SLAM system built on ORB-SLAM2, adding a dedicated semantic segmentation thread using Light-Weight RefineNet trained on PASCAL VOC2012. Instead of segmenting every frame—a computational bottleneck—the algorithm adaptively decides when segmentation is needed based on feature-point motion levels, enabling cross-frame semantic information transfer.
Each feature point is assigned a motion level (0–10) and classified as static, movable static, or dynamic. Static points drive initial pose estimation, followed by a second optimization using weighted static constraints. For frames without segmentation, motion levels are propagated from matched points in previous frames, keyframes, or local map points, with fallback strategies for unmatched points.
Tests on the TUM RGB-D dynamic scene dataset showed the algorithm outperformed ORB-SLAM2 and other dynamic-scene SLAM methods on most sequences, improving localization accuracy by up to 90.57% in indoor dynamic environments.
Why it mattered
This work demonstrated that integrating deep semantic priors with adaptive computation can make SLAM both accurate and practical for real-time use—an important step for robots operating alongside humans in factories, warehouses, and other dynamic industrial settings.
“The positioning accuracy in indoor dynamic environments can be improved by 90.57% compared with the ORB-SLAM algorithm without the improvement.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2023-03-06 · “动态场景下基于自适应语义分割的RGB-D SLAM算法”