SemanticLoop: Loop Closure With 3D Semantic Graph Matching

Loop closure can effectively correct the accumulated error in robot localization, which plays a critical role long-term navigation of robot. Traditional appearance-based methods rely on local features and are prone to failure ambiguous environments. On other hand, object recognition infer objects' category, pose, extent. These objects serve as stable semantic landmarks for viewpoint-independent non-ambiguous loop closure. However, there is object-level data association problem due lack efficient robust algorithms. We introduce novel algorithm, incorporates IoU, instance-level embedding, detection uncertainty, formulated linear assignment problem. Then, we model TSDF volumes represent environment 3D graph with semantics topology. Next, propose matching-based based reconstructed graphs by aligning matched objects. Finally, refine poses camera trajectory an pose optimization. Experimental results show that proposed method significantly outperforms commonly used nearest neighbor accuracy. Our more environmental appearance changes than existing methods.