Probabilistic Robot Localization using Visual Landmarks

TL;DR: This project examined the feasibility of using the probabilistic Monte Carlo localization algorithm to estimate a robot’s location based off of occasional visual landmark cues and suggested that with minor modifications the system could become a reliable localization scheme.

Abstract: Effective robot navigation and route planning is impossible unless the position of the robot within its environment is known. Motion sensors that track the relative movement of a robot are inherently unreliable, so it is necessary to use cues from the external environment to periodically localize the robot. There are many methods for accomplishing this, most of which either probabilistically estimate the robot’s movement based on range sensors, or require having enough unique visual landmarks present to geometrically calculate the robot’s position at any time. In this project I examined the feasibility of using the probabilistic Monte Carlo localization algorithm to estimate a robot’s location based off of occasional visual landmark cues. Using visual landmarks has several advantages over using range sensor data in that landmark readings are less affected by unexpected objects and can be used for fast global localization. To test this system I designed a robot capable of navigating Olin-Rice by observing pieces of colored paper placed at regular intervals along the halls as an extension of my summer 2005 research on RUPART. The localization system could not localize the robot in many situations due to the sparse nature of the landmarks, but results suggest that with minor modifications the system could become a reliable localization scheme.