Filtering methods like the Extended Kalman Filter fail due to quadratic complexity in state size. Graph SLAM overcomes this by modeling the trajectory and environment as a sparse graph of poses and constraints. We detail the non-linear least squares formulation, where the objective is minimizing the sum of weighted squared errors. Learn how odometry constraints define local structure and how the information matrix dictates measurement certainty. Crucially, we analyze loop closure: the mechanism for detecting revisited locations and integrating non-local edges to enforce global metric consistency. Finally, explore the role of sparse matrix solvers and frameworks like General Graph Optimization and Ceres Solver in achieving scalable, globally optimized localization.
00:00: Filtering Limits and Graph Paradigm
00:49: Pose Graph Nodes and Edges
01:29: Minimizing Global Cost Function
02:17: Odometry and Prior Constraints
02:59: Loop Closure for Global Consistency
03:37: Robust Place Recognition Techniques
04:17: Adding the Non-Local Edge
04:55: Iterative Non-Linear Optimization
05:32: G2O and Ceres Solver Frameworks
06:17: Consistency, Batching, and ISAM
#Robotics #SLAM #GraphSLAM #LoopClosure #PoseGraph #Optimization #G2O #CeresSolver
