Evolving belief network for resource-efficient place recognition in unknown environments / von MS-CSE Syed Muhammad Ali Musa Kazmi; erster Gutachter: Prof. Dr. Erdal Kayacan, zweiter Gutachter: Prof. Dr. Sybille Hellebrand. Paderborn, 2025
Content
- Dedication
- Declaration
- Abstract
- Zusammenfassung
- Acknowledgement
- Contents
- 1 Introduction
- 2 Background
- 2.1 Simultaneous Localization & Mapping (SLAM)
- 2.2 Essentials of Vision-based SLAM
- 2.3 Representing the Observed World
- 2.4 Classic Map Estimation Algorithms
- 2.5 Place Recognition
- 2.5.1 Defintion of a Place
- 2.5.2 Emergence of Appearance-based Mapping
- 2.5.3 Data Association – in a Metric Map vs. Image Space
- 2.5.4 Evaluation Criteria
- 2.6 Visual Description of a Place
- 2.6.1 Intensity-based Descriptor
- 2.6.2 Vocabulary-based Descriptor
- 2.6.3 Gist Descriptor
- 2.6.4 HOG Features
- 2.6.5 CNN-based Descriptor
- 2.7 Biology of Visual Perception, Learning and Navigation
- 3 Related Works
- 3.1 Vocabulary-based Place Recognition
- 3.2 Matching Sequences for Place Recognition
- 3.3 Learning Spatial Representation for Place Recognition
- 3.4 Bio-inspired Paradigms
- 3.5 CNN-based Place Recognition
- 4 Bio-inspired Framework for On-the-fly Visual Learning
- 4.1 RatSLAM System: A Bird's-eye View
- 4.2 Proposed Model for Visual Learning
- 4.3 Leveraging Proposed Model for Place Recognition
- 4.4 Evaluation and Results
- 4.4.1 Experimental Setup
- 4.4.2 Evolution of View Cells: Proposed vs. Baseline Model
- 4.4.3 Robustness to Sensory Perturbations
- 4.4.4 Precision–Recall Scores
- 4.5 Summary
- 5 Growing Belief Network for Real-time Place Recognition
- 5.1 Introduction
- 5.2 Novel Network Dynamics for Visual Learning
- 5.3 Probabilistic Modeling of the Belief Network
- 5.4 Evaluation and Results
- 5.4.1 Experimental Setup
- 5.4.2 Correctness of the Learned Representation
- 5.4.3 Evolution of Neurons in the Network
- 5.4.4 Loop Detection Accuracy
- 5.4.5 Discussion
- 5.5 Summary
- 6 Expectancy Detection of a Place through Nearby Context
- 6.1 Introduction
- 6.2 Capturing the Spatial Layout of Scenes
- 6.3 Improved Learning Dynamics for GSOM
- 6.4 Loop-closure Detection
- 6.5 Evaluation and Results
- 6.5.1 Experimental Setup
- 6.5.2 Loop Detection Accuracy
- 6.5.3 Comparison with the Baseline Techniques
- 6.5.4 Time Requirements
- 6.5.5 Discussion on System's Parameters
- 6.5.6 On the Effect of Weight Normalization
- 6.5.7 Correctness of the Learned Representation
- 6.5.8 Evolution of Neurons in the Network
- 6.5.9 On Extensibility to Other Feature Spaces
- 6.6 Summary
- 7 Conclusion and Future Research
- 7.1 Thesis Summary
- 7.2 Biological Fidelity in a Technical System
- 7.3 This Research vs. State-of-the-Art
- 7.4 Future Research Directions
- 7.4.1 Search Optimization
- 7.4.2 Extension of the Place Recognition Subsystem
- 7.4.3 On Learning the Geometric Maps
- 7.4.4 Optimality of the Hyperparameters
- 7.4.5 Then GSOM Meets the CNNs
- 7.4.6 Other Considerations
- 7.5 Concluding Remarks
- Bibliography
- List of Publications
- List of Abbreviations
- List of Notations
- List of Figures
- List of Tables