A-SPAM: A Novel Asynchronous Semantic Padding and Matching Integrated Framework for Dynamic Loop Closure Detection
QiBin He, Yapeng Wang, Yanming CHAI, Qiyue Huang, Tiankui Zhang, KeiIm Sio, Jie Zhang
AI summary
Problem
Traditional visual loop closure detection methods struggle in dynamic environments where moving objects dominate the scene, causing sparse static features, occlusions, and unreliable feature matching that lead to catastrophic map corruption or drift.
Approach
The framework asynchronously tracks static semantic entities across frames to build a spatiotemporal graph, then validates loop closures by matching the graph's rigid topology and visual features using a Siamese network.
Key results
- ≥76.8% recall at 100% precision on dynamic TUM/BONN sequences
- <0.07 m mean translational error under degraded odometry
- Cross-frame semantic consistency algorithm eliminates trajectory redundancy
- Rigid topology constraint matching enables robust loop validation despite occlusions
Why it matters
Enables reliable autonomous navigation for robots and drones in crowded, dynamic settings where conventional SLAM systems typically fail.
Abstract
Loop closure detection in dynamic SLAM faces crit- ical challenges when dynamic objects dominate camera views, degrading frame-to-frame methods reliant on static landmarks. We proposeA-SPAM,anasynchronousframeworkthatconstructsspa- tiotemporal semantic graphs via semantic padding (entity track- ing + rigid structure analysis) and validates loops via seman- tic matching (topology-feature hybrid correlation). Evaluated on TUM and BONN datasets, A-SPAM achieves at least 76.8% recall rate at 100% precision in dynamic environments, while main- taining a mean translational error of less than 0.07 m across dynamic sequences under degraded odometry conditions. The pro- posed framework corrects erroneous trajectories and enhances robustness against odometry failures in dynamic environments.