A semantic gaussian-spraying dynamic RGB-DSLAM method and system with loop closure optimization
By employing semantic segmentation and loop closure optimization methods, the problems of error accumulation and pose-map inconsistency in 3D Gaussian sputtering SLAM in dynamic scenes were solved, achieving high-precision static scene reconstruction and global consistency, and improving the system's tracking stability and mapping accuracy.
Patent Information
- Application Number
- CN202610845707.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-12
- Publication Date
- 2026-07-17
AI Technical Summary
Existing 3D Gaussian sputtering SLAM methods perform poorly in dynamic scenes. The lack of online loop closure correction leads to error accumulation and map drift. Furthermore, traditional loop closure optimization frameworks fail to adapt to dense representation characteristics, resulting in pose-map inconsistency and reduced mapping accuracy.
A semantic segmentation model is used to extract dynamic object masks, and a bag-of-words model is used to retrieve loop closure candidate frames. Geometric registration is performed through an iterative nearest point algorithm, and global calibration is completed using pose graph optimization. The optimal pose is fixed for full posterior reconstruction, dynamic Gaussian points are pruned, and a high-precision static scene map is constructed.
It significantly improves the tracking stability and scene reconstruction consistency of the SLAM system in dynamic scenes, effectively suppresses dynamic target interference, improves localization and mapping accuracy, and reduces pose error.
Smart Images
Figure CN122415741A_ABST