蒋小强, 卢虎, 闵欢. 基于连续-离散MRF图模型的鲁棒多机器人地图融合方法[J]. 机器人, 2020, 42(1): 49-59. DOI: 10.13973/j.cnki.robot.190329
引用本文: 蒋小强, 卢虎, 闵欢. 基于连续-离散MRF图模型的鲁棒多机器人地图融合方法[J]. 机器人, 2020, 42(1): 49-59. DOI: 10.13973/j.cnki.robot.190329
JIANG Xiaoqiang, LU Hu, MIN Huan. Robust Multi-robot Map Merging Based on Continuous-Discrete MRF Graph Model[J]. ROBOT, 2020, 42(1): 49-59. DOI: 10.13973/j.cnki.robot.190329
Citation: JIANG Xiaoqiang, LU Hu, MIN Huan. Robust Multi-robot Map Merging Based on Continuous-Discrete MRF Graph Model[J]. ROBOT, 2020, 42(1): 49-59. DOI: 10.13973/j.cnki.robot.190329

基于连续-离散MRF图模型的鲁棒多机器人地图融合方法

Robust Multi-robot Map Merging Based on Continuous-Discrete MRF Graph Model

  • 摘要: 针对多机器人同步定位与建图(MSLAM)中感知偏差会产生高度相关且互一致的异常回环,进而导致定位与地图变形等问题,提出了基于马尔可夫随机场(MRF)的通用连续-离散图模型.其中,连续图对标准位姿图(pose graph)进行建模;离散图通过对异常值相关关系的显式建模,建立剔除模型.在此基础上,进一步利用凸松弛方法,将连续-离散图代表的非凸且NP(非确定性多项式)完全的组合优化问题转化为半正定规划(SDP)问题,方便利用现有凸优化工具进行求解.仿真和实测数据实验表明,本文方法提高了位姿图对感知偏差带来异常外点的鲁棒性,且结果不依赖于位姿初始值的好坏,在异常值占比为50%的情况下,剔除率仍可达99.8%,地图融合精度优于现有主流动态协方差缩放(DCS)方法和两两一致测量集(PCM)方法.

     

    Abstract: In the multi-robot simultaneous localization and mapping (MSLAM), perceptual aliasing will generate highly correlated and mutual-consistent spurious inter-map loop closures, which will lead to the failure of location and map deformations. A general continuous-discrete graph model based on Markov random field (MRF) is proposed to solve this problem. The continuous graph models the standard pose graph, and the discrete graph establishes the rejection model by explicit modeling of the correlation of outliers. On this basis, the convex relaxation method is further used to transform the non-convex and NP (non-deterministic polynomial) complete combinatorial optimization problem represented by the continuous-discrete graph into a semidefinite programming (SDP) problem, which is convenient to solve by using off-the-shelf convex solvers. Simulation and experiment results demonstrate that the proposed method can improve the robustness of the pose graph to the outliers caused by perceptual aliasing, and the result doesn't depend on the initial value of the pose. In the case of an outlier ratio of 50%, the rejection rate is still up to 99.8%, and the map fusion accuracy is better than the existing state-of-the-art DCS (dynamic covariance scaling) and PCM (pairwise consistent measurement) methods.

     

/

返回文章
返回