TL;DR
This paper presents a unified approach to loop closing and recovery in real-time monocular SLAM, treating both as graph edge creation. By integrating appearance and structure-based matching within a graph framework, their system achieves robust localization and mapping in dynamic environments, even after tracking failures.
What Problem Does This Paper Address?
Simultaneous Localization and Mapping (SLAM) systems, particularly those using a single camera (monocular SLAM), face significant challenges in maintaining accurate and consistent maps over time. A primary issue is "tracking failure," where the system loses track of the camera's position due to factors like blur, occlusion, or rapid motion. This can lead to catastrophic corruption of the map. Furthermore, even when tracking is successful, maps can become globally incoherent, especially over large scales, and "loop closing" – recognizing that the camera has returned to a previously visited location – is difficult to achieve reliably in real-time. Existing methods often treat recovery from failure and loop closing as separate, complex problems, leading to fragile systems that are unusable for many real-world applications.
Methods
The core of this work is a unified approach to loop closing and recovery within a graph-based SLAM framework, building upon the Graph-SLAM system by Eade and Drummond.
- Unified Graph Representation: Both loop closing and recovery are framed as the creation of a new edge in the SLAM graph.
- Loop Closing: Creates an edge between two nodes within the same connected component, adding a constraint that improves global map coherence.
- Recovery: Creates an edge between nodes in different connected components, effectively merging them and allowing mapping to continue seamlessly after a tracking failure.
- Appearance-Based Candidate Selection: To efficiently find potential loop closure or recovery candidates, the system uses a bag-of-words (BoW) appearance model.
- Online Vocabulary Building: Instead of relying on pre-trained vocabularies, the system builds its visual word vocabulary incrementally during operation. This involves maintaining a main vocabulary and a "young word" database for new candidates.
- Appearance Model: For each graph node, the system stores a histogram of observed visual words. A term-frequency-inverse-document-frequency (TF-IDF) scheme is used to rank nodes based on their expressed words relative to the current image's observed words, identifying the top k candidate nodes.
- Structure and Appearance Matching: Once candidate nodes are identified, a more robust matching process is employed:
- Local Landmark Appearance Model: Each graph node maintains a local appearance model of its landmarks, distinct from the global BoW model.
- Descriptor Matching: Interest points from the current image are matched against the local appearance models of candidate nodes.
- Robust Model Fitting (MLESAC): A robust estimator (MLESAC) is used to find the most consistent correspondences between image features and candidate node landmarks, from which a pose estimate is derived.
- Candidate Edge Trial Period: A newly established candidate edge undergoes a "trial period" where its viability is scored based on the success rate of landmark predictions. If successful predictions outweigh failures, the edge is validated and becomes a permanent graph edge, triggering loop closure or recovery.
- Basic Modifications: The system uses a scale-space extrema detector for interest points, providing scale information. It also allows for multiple connected components in the graph, crucial for handling tracking failures by starting a new component.
Datasets
Not explicitly discussed in the paper. The authors mention using "indoor and outdoor sequences" and "completely planar real scene" and "synthetic sequences" for testing, but no specific named datasets are provided.
Key Insights and Ideas
- Unified Framework: The most significant insight is treating loop closing and recovery as fundamentally the same graph operation (edge creation). This simplifies the system architecture and allows for seamless transitions between normal operation, failure, and re-localization.
- Continuous Mapping: By creating new graph components upon failure, the system ensures that mapping never stops. This is a major advantage over methods that halt mapping or require extensive relocalization before resuming.
- Online Vocabulary Building: The ability to build the visual vocabulary online is crucial for real-time applications and adaptability to new environments without requiring extensive offline training.
- Appearance-Structure Synergy: The combination of a coarse, global BoW model for candidate selection and a fine-grained local appearance model for robust matching is key to efficiently and accurately identifying loop closure/recovery opportunities.
- Recovery as Component Merging: Framing recovery as merging disconnected graph components elegantly handles the transition from a lost state back into the main map, preserving previously mapped information.
Key Results
- Real-Time Performance: The system achieves per-frame processing under 33 ms and loop detection/recovery in under 6 ms on a 2.2 GHz dual-core computer.
- Robustness: The system successfully closes loops and recovers from tracking failures in both indoor and outdoor sequences.
- Scalability: It can map thousands of landmarks (e.g., 251 in a planar scene, 1043 in an outdoor loop, 1402 in an indoor scene) while operating in real-time.
- Accuracy: In a planar scene test, landmarks with low depth uncertainty were accurately placed relative to the maximum likelihood plane.
Limitations and Future Work
- Spurious Loop Closure: The system can be susceptible to spurious loop closures in environments with extensive, repeated structures (especially in synthetic sequences). The authors suggest a probabilistic model for appearance-based loop closure could mitigate this.
- Cycle Optimization: The current cycle optimization treats edge transformations as independent, which can lead to over-confident and incorrect global maps when many loops are optimized. Future work will involve addressing this through conservative local graph optimizations.
- Larger Environments: The authors intend to evaluate the performance and discrimination of their bag-of-words model in much larger environments where the graph size is significantly larger.
Summary updated