联合残差网络和最近点迭代的机器人重定位

Robot relocation based on residual networks and iterative closest point

针对机器人遭遇绑架、系统故障重启而产生的定位丢失问题,提出一种基于ResNet的机器人重定位方法。所提方法将重定位分为基于残差网络(residual network,ResNet)的粗匹配和基于最近点迭代(iterative closest point,ICP)细匹配2个阶段。在粗匹配阶段,将激光点云数据转换为图像,然后将相邻时间的图像堆叠成多通道图像作为ResNet的输入,以增强图像的时序特征。在细匹配阶段,ResNet输出机器人的预测位置,并将预测结果作为ICP算法的初值进行点云细匹配,从而获取最终位姿。对于相似环境,提出动态重定位方法,通过移动机器人进行多次重定位避免误匹配的情况。仿真实验结果表明:该方法与增强蒙特卡罗定位(augmented Monte Carlo localization,AMCL)算法进行了对比,定位用时降低了8.2s,定位成功率提升了43.4%,证明了该算法具有更好的重定位效果。

A mobile robot relocation algorithm based on ResNet was proposed to solve the problem of positioning loss caused by robot kidnapping and system failure restart.The proposed algorithm divides relocation into two stages:coarse matching based on ResNet and fine matching based on iterative closest point(ICP).In the coarse matching stage,the laser point cloud data is first converted into images,and then adjacent images in time are stacked into multi-channel images as input to ResNet to enhance the temporal features of the images.In the fine matching stage,the predicted position of the robot output by ResNet is used as the initial value of the ICP algorithm for point cloud fine matching,in order to obtain the final pose.For similar environments,a dynamic relocation method was proposed,which uses mobile robots to perform multiple relocations to avoid mismatches.Finally,a comparative experiment was conducted with the relocation algorithm augmented Monte Carlo localization(AMCL)in the simulation experiment.The localization time was reduced by 8.2 s and the localization success rate was improved by 43.4%,proving that the proposed algorithm has better relocation performance.