Researchers at Shanghai University of Engineering Science have developed a real-time visual SLAM algorithm designed for dynamic indoor environments. The method integrates semantic segmentation, motion blur detection, and adaptive-threshold epipolar constraints to effectively remove dynamic feature points, addressing challenges posed by moving objects and camera motion blur.
Key takeaways
- Combines YOLACT semantic segmentation with motion blur detection and adaptive epipolar constraints to identify dynamic objects in prior regions.
- Uses K-means clustering on superpixels to detect non-prior dynamic objects, improving robustness without heavy computational cost.
- On the TUM RGB-D dataset, the algorithm reduces absolute trajectory error (ATE) RMSE by at least 94.26% compared to ORB-SLAM2, with relative pose error (RPE) RMSE reduced by at least 87.95% (translation) and 92.07% (rotation).
The approach is built on ORB-SLAM2, modifying the tracking thread to include semantic segmentation and dedicated modules for prior and non-prior regions. Motion blur is assessed using a combination of standard deviation and local entropy, while an adaptive threshold based on feature point velocity and depth overcomes limitations of fixed thresholds in varying depth and speed conditions. This work offers a practical solution for mobile robots and autonomous systems operating in cluttered, dynamic settings.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-05-12 · “结合运动模糊检测与自适应极线约束的动态V-SLAM算法”