What this project is
MonoSLAM is a monocular visual SLAM build that tracks camera motion, estimates sparse structure, and incrementally refines the map in real time.
Why it matters
Monocular SLAM sits at the boundary between geometry and systems engineering. It is a useful testbed for understanding uncertainty, observability, and failure modes in perception pipelines.
Architecture overview
- Feature extraction and matching in consecutive frames.
- Two-view bootstrap with essential matrix, cheirality checks, and parallax constraints.
- Keyframe insertion and pose-only optimisation.
- Local map updates and outlier rejection.
Current status
- Front-end feature matching implemented with reproducible parameter settings.
- Two-view bootstrap working on controlled sequence subsets.
- Keyframe tracking integrated; map consistency checks in progress.
Selected results
- Stable bootstrap on short ETH3D sequences with robust inlier filtering.
- Tracking retained through moderate viewpoint changes.
- See the demo pages for visual output snapshots and diagnostics.
Roadmap
- Add loop-closure candidates and relocalisation hooks.
- Improve failure detection around low-parallax segments.
- Expand quantitative benchmarking across more sequence types.