loop closure detection

The semantic visual SLAM field lacks a systematic survey, particularly regarding the integration of deep learning and large language models (LLMs). To address this gap, we propose a unified problem formulation that decomposes semantic SLAM into five core modules: visual localization, semantic feature extraction, map construction, data association, and loop closure optimization. We introduce a modular analytical framework that unifies classical geometric approaches with modern semantic understanding techniques—including semantic segmentation, object detection, scene understanding, and LLM-based reasoning—and conduct empirical evaluations on benchmark datasets. Our work provides the first comprehensive taxonomy of technical evolution, critically analyzes limitations of existing methods, and identifies key bottlenecks: semantic consistency, cross-modal alignment, and real-time performance. The study establishes an authoritative knowledge base and a scalable technical roadmap for future research in semantic SLAM.

Current semantic SLAM systems face limitations in semantic understanding depth, robustness of loop closure detection, and stability of data association. This work proposes an RGB-based semantic SLAM framework grounded in a unified geometry-instance representation, which for the first time extends geometric foundation models to jointly predict dense geometric structures and view-consistent instance embeddings. By introducing viewpoint-invariant semantic anchors, the method bridges the gap between geometric reconstruction and open-vocabulary semantic mapping, enabling semantic-coherent data association and instance-guided loop closure detection. Experimental results demonstrate that the proposed system significantly outperforms state-of-the-art approaches in terms of map consistency and reliability under large-baseline loop closure scenarios.

In multi-map SLAM, submap disconnection due to robot localization failure impedes robust submap registration. To address this, we propose a cross-submap robust registration method leveraging visual place recognition (VPR) as a semantic bridge. Unlike conventional VPR-based loop closure detection, our approach elevates VPR to a structured geometric prior for submap association—enabling alignment across non-overlapping domains, disparate time periods, and heterogeneous sensor modalities. The method integrates deep feature matching, geometric consistency verification, and graph-based pose optimization, incorporating a lightweight VPR retriever and a differentiable pose solver. Evaluated on KITTI, NCLT, and a custom multi-submap dataset, our method achieves a 23.6% improvement in submap connection success rate, reduces absolute pose error to 0.48 m / 1.2°, and operates at 12 fps.

This work addresses the degradation of multi-session mapping accuracy in autonomous systems operating in repeatedly visited environments, where accumulated drift and map inconsistency pose significant challenges. To this end, the authors propose a topology-aware incremental mapping and localization framework that analyzes the pose graph’s topological structure to identify poorly connected regions. Leveraging an uncertainty-driven mechanism, the system dynamically triggers selective loop closure detection and map fusion, eschewing conventional full SLAM post-alignment strategies. This approach enables efficient and globally consistent map construction across sessions. Extensive experiments on publicly available overlapping sequences and real-world mine tunnel datasets demonstrate the method’s effectiveness, showing a significant reduction in cumulative error and marked improvement in global consistency of multi-session maps.

To address key bottlenecks in large-scale LiDAR-based navigation—namely, high computational cost, sensitivity to viewpoint changes, and poor robustness in long-range loop closure detection—this paper proposes a subgraph consistency–based relocalization framework. Methodologically, it introduces geometric consistency verification, rather than embedding distance, as the primary loop closure criterion—a novel departure from conventional approaches. The framework jointly exploits rotation-invariant object features, object-level semantics, and graph neural networks (GNNs) to model local neighborhood structure, while employing a bag-of-words (BoW)–style global subgraph feature matching for efficient cross-view retrieval. Experimental results demonstrate that the method achieves state-of-the-art (SOTA) accuracy while doubling inference speed, significantly improving both efficiency and robustness for long-range and multi-view loop closure detection in large-scale environments.

This work proposes a memory- and speed-efficient online semantic SLAM framework to address the challenge of simultaneously achieving geometric consistency, semantic coherence, and computational efficiency in indoor 3D semantic mapping and navigation. By integrating SLAM with semantic mapping, the method constructs a spatiotemporally consistent 3D scene understanding system based on a VGGT tracking head. It processes arbitrarily long video streams via a sliding window, aligns local submaps using camera poses, and lifts 2D instance masks into 3D objects with consistent identity labels. Floor-plane projection is further leveraged to enable efficient navigation support. Evaluated on ScanNet++, the approach achieves competitive point cloud performance while consuming less than 17 GB of GPU memory, enabling real-time interactive assistive navigation with audio feedback.

This work addresses the challenges of cross-view geolocalization in unseen regions and complex conditions, where viewpoint variations and domain shifts lead to semantic inconsistency and limited generalization. To overcome these issues, the authors propose GeoLink, a framework that leverages offline multi-view image reconstruction to generate scene point clouds, thereby establishing a 3D structural prior. GeoLink integrates this 3D geometry into 2D feature learning through a geometric-aware semantic refinement module and a unified view relation distillation mechanism. This design enhances semantic consistency while preserving the computational efficiency of purely 2D inference. Extensive experiments demonstrate that GeoLink significantly outperforms existing methods across multiple benchmark datasets, exhibiting superior cross-domain generalization under diverse weather conditions and in previously unseen domains.

This work addresses the challenge of cross-view localization for robots operating in unseen urban environments using publicly available semantic maps such as OpenStreetMap. The proposed method leverages a vision-language model (VLM) to extract semantic landmarks from panoramic images and employs a lightweight matcher—distilled from the VLM via knowledge distillation—to efficiently align these observations with large-scale semantic maps. Temporal pose estimates are refined through Bayesian filtering. Trained solely on daytime data from a single city, the framework demonstrates robust generalization across eleven diverse environmental conditions, including nighttime and blizzard scenarios, and scales to map areas spanning hundreds of square kilometers. The authors also release a semantic dataset and source code to support reproducibility and further research.

This work addresses the challenge of pose ambiguity in indoor environments with repetitive or symmetric layouts, where conventional graph-matching approaches relying solely on geometric structure often fail. To overcome this limitation, the authors propose a semantic-enhanced graph matching method that explicitly models semantic relationships between RGB-D detected objects and structural elements such as rooms and walls. These semantic cues are leveraged to pre-filter matching candidates, which are subsequently refined through geometric verification to improve SLAM robustness. Integrated into the iS-Graphs framework, the approach constructs scene graphs and performs semantic relation reasoning, substantially reducing the number of candidate matches in both synthetic and simulated environments. This leads to notable gains in computational efficiency and convergence speed, with clear performance advantages over purely geometric methods in symmetric scenes.

Existing semantic 3D Gaussian SLAM methods are prone to persistent semantic artifacts in complex scenes due to unreliable 2D semantic priors, manifesting as mislabeled occlusions, boundary leakage, and premature labeling of ambiguous regions. This work proposes a geometry-validated semantic evidence fusion framework that introduces, for the first time, a geometric reliability assessment mechanism. By dynamically evaluating the credibility of semantic observations through visibility consistency, surface support constraints, and ray-wise conflict uncertainty, the method adaptively adjusts optimization weights during inference. This approach effectively suppresses erroneous updates in occluded regions, mitigates geometry-unsupported semantic diffusion, and defers label assignment in ambiguous areas. Experiments demonstrate significant improvements on Replica in semantic consistency, boundary fidelity, and reconstruction quality, while maintaining tracking accuracy on ScanNet comparable to state-of-the-art methods.

This work addresses the challenge of robust drone localization in GNSS-denied environments, where visual-inertial odometry suffers from cumulative drift and appearance-based methods are sensitive to seasonal changes, lighting variations, viewpoint shifts, and map obsolescence. To overcome these limitations, the authors propose a semantic map localization framework that integrates structured semantic geometry with a stability-aware mechanism, leveraging persistent environmental structures such as roads and buildings. The approach employs semantic grid alignment, relational graph matching, geospatial saliency evaluation, and a ternary observation model (positive, contradictory, or unknown), complemented by an integrity-aware fuzzy rejection strategy. Experimental results demonstrate a Recall@1 of 94.5–95.5% across 220 cross-view queries, substantially outperforming global semantic descriptors (58.6%) and confirming the method’s robustness to perturbations including rotation, scaling, and occlusion.