Information gain-based SLAM graph pruning

In graph-based simultaneous localization and mapping, the dimension of nonlinear constraint equations increases linearly with the distance and duration of robots motion. An efficient approach based on information gain is proposed to prune the graph. By evaluating the relative variation of features' information matrices before and after the pruning, any observation information below the given threshold of the robot pose is pruned, as well as corresponding observations, so that the complexity of SLAM optimization problem is simplified significantly. Exact and approximate computation methods of information gain are provided, according to the assumption of spherical covariance of measurements. The connectivity of the pruned graph is kept using the recovered pruning method. Experimental results based on Monte Carlo simulation and opensource environment dataset show that: around 90% of poses and features are pruned, on the premise that the optimization errors are not introduced apparently. The optimization efficiency is raised greatly.