visual slam

**Visual SLAM (vSLAM)** is the **SLAM specialization that relies primarily on camera imagery to estimate trajectory and build maps** - it spans monocular, stereo, and RGB-D setups with different tradeoffs in scale observability and robustness. **What Is Visual SLAM?** - **Definition**: Camera-driven SLAM pipeline combining visual odometry, mapping, and loop closure. - **Sensor Variants**: Monocular, stereo, and RGB-D each provide different geometric constraints. - **Map Types**: Sparse landmark maps, semi-dense maps, or dense reconstructions. - **Runtime Goal**: Real-time pose tracking with persistent environment model. **Why Visual SLAM Matters** - **Low-Cost Hardware**: Cameras are inexpensive and widely available. - **Rich Semantics**: Visual features support object-aware mapping and scene understanding. - **Indoor and AR Strength**: Core technology for headsets and mobile mapping. - **Scalability**: Works from small rooms to large outdoor routes with proper design. - **Research Maturity**: Strong ecosystem of feature-based and direct methods. **vSLAM Building Blocks** **Tracking Front-End**: - Match visual features between frames. - Estimate camera motion relative to map. **Local Mapping**: - Triangulate landmarks and refine nearby poses. - Maintain keyframe graph. **Loop Closure and Global BA**: - Detect revisited places and correct drift. - Optimize global consistency. **How It Works** **Step 1**: - Track camera with frame-to-map matching and estimate incremental pose. **Step 2**: - Update map and periodically run loop-closure-driven global optimization. Visual SLAM is **a camera-centric localization and mapping framework that balances affordability, semantic richness, and real-time performance** - it is one of the most deployed SLAM paradigms in robotics and AR.