基于信息增益的SLAM图精简  

Information Gain-based SLAM Graph Pruning

作　　者：程见童[1] 江振宇[1] 张银辉[1] 张为华 

机构地区：[1]国防科学技术大学航天科学与工程学院,湖南长沙410072

出　　处：《机器人》2014年第5期527-534,共8页Robot

基　　金：国家自然科学基金资助项目(11102229)

摘　　要：针对基于图论的同时定位与制图中,非线性约束方程组维数随机器人运行距离和时间的增加而不断增大的问题,提出一种基于信息增益的图精简算法.该算法通过评估精简前后特征点信息矩阵相对变化,删除观测信息量小于给定阈值的机器人位姿及相应的观测,达到显著简化优化问题的目的.根据测量球形协方差矩阵假设,给出了信息增益的精确和近似计算方法.通过恢复性图剪枝方法,确保图精简过程中的连通性.蒙特卡洛仿真和开源实验数据计算结果表明,在不引入明显的优化误差前提下,该方法可实现位姿和特征点90%的精简,显著提高图优化效率.In graph-based simultaneous localization and mapping, the dimension of nonlinear constraint equations increas- es 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 ob- servation 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.

关 键 词：同时定位与制图 信息增益 图优化 图精简 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]