Local Accuracy in Global Uncertainty

Abstract

Mobile robots need to interact with their environment to perform their tasks. To be successful they often need to know what their surrounding looks like, and where they are located in that surrounding. The act of simultaneously estimating both the state of the robot and the state of the environment is called Simultaneous Localization And Mapping (SLAM). The SLAM problem has been intensively researched since it is one of the key prerequisites for correct functioning of mobile robots. Popular mapping algorithms use Kalman or particle filters to describe the robot state and the environment with probability distributions. When the mapped environment becomes larger these algorithms need more computational resources and have trouble recognizing known locations. This is mostly because the stored map becomes too large to process and the uncertainties have become too big to properly be described or used in the SLAM framework. This work proposes a Hybrid Metric-Topological (HMT) mapping method to solve problems regarding place recognition and computational resources commonly experienced when mapping large environments. The framework will create local maps using a particle filter and connect the local maps in a graph constrained with edges. The local environment is mapped accurately, while the relative uncertainties are stored in the edges. In this work the devised framework is called "Forced Resampling hybrid Metric-Topological Mapping" or FoRMeT Mapping. It is characterized by being an approximate solution to the SLAM problem that makes justifiable assumptions to be a lightweight HMT framework. The biggest innovation is the forced resampling strategy which allows the particle filter ambiguity to be stored in the last two visited local maps, while keeping all other maps in a shared topology. An implementation of the framework is devised in C++ with ROS (Robot Operating System) and tested on a real world environment containing a large loop, and a simulated large scale environment. The same experiments are also performed with the commonly acknowledged mapping algorithm "GMapping" for comparison. FoRMeT Mapping outperforms GMapping on all performance criteria. FoRMeT uses about 3 times less computational resources for each task and the local areas look as good as GMapping’s local areas, with exceptions around the loop closure areas where FoRMeT looks better with less discrepancies. It can be concluded that FoRMeT can close large loops, map large environments and outperform GMapping while remaining a lightweight SLAM framework.