Metrics
1 Record Views
Book chapter
Loop Closure Detection Revisited: A Clustering Perspective
Applied Imagery Pattern Recognition Workshop, pp.415-429
Lecture Notes in Computer Science, vol 16446, Springer
Applied Imagery Pattern Recognition Workshop, AIPR 2025, 53rd (Washington, District of Columbia, USA, 10/13/2025–10/14/2025)
04/01/2026
Abstract
Loop closure detection (LCD) is critical for reducing drift and maintaining map consistency in SLAM systems, yet image-retrieval–based methods struggle with perceptual aliasing, viewpoint and appearance changes, and limited scalability. We reformulate LCD as clustering in a learned latent space rather than database retrieval. A convolutional autoencoder (CAE) is first pre-trained on environment imagery to produce compact, structure-aware embeddings. We then create globally-aware descriptors with a new model. During operation, keyframes are encoded and compared in latent space against a growing memory set. If an embedding lies beyond a threshold distance, it is considered a potential loop closure and added to the clustering structure. To enforce spatially meaningful structure, we apply triplet loss: the immediate previous frame serves as a positive (temporal proximity), while other keyframes act as negatives, encouraging embeddings from the same place to cluster and distinct places to separate. This design improves robustness to aliasing and appearance variation and reduces computational cost by avoiding exhaustive database search. Experiments on multiple place-recognition and navigation datasets show competitive or superior precision–recall performance compared to NetVLAD, DBoW2, and AP-GeM with more temporally consistent loop-closure clusters. Overall, the results indicate that latent-space clustering with globally-aware descriptors provides a scalable and robust alternative to conventional retrieval-based LCD.
Details
- Title
- Loop Closure Detection Revisited: A Clustering Perspective
- Publication Details
- Applied Imagery Pattern Recognition Workshop, pp.415-429
- Resource Type
- Book chapter
- Conference
- Applied Imagery Pattern Recognition Workshop, AIPR 2025, 53rd (Washington, District of Columbia, USA, 10/13/2025–10/14/2025)
- Publisher
- Springer; Cham
- Series
- Lecture Notes in Computer Science; vol 16446
- Format
- supplemental link
- Copyright
- The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2026
- Identifiers
- 99381730223006600
- Academic Unit
- Intelligent Systems and Robotics; Hal Marcus College of Science and Engineering
- Language
- English