Indoor mobile robot localization in dynamic and cluttered environments using artificial landmarks

Abstract

Purpose Robot localization in dynamic, cluttered environments is a challenging problem because it is impractical to have enough knowledge to be able to accurately model the robot’s environment in such a manner. This study aims to develop a novel probabilistic method equipped with function approximation techniques which is able to appropriately model the data distribution in Markov localization by using the maximum statistical power, thereby making a sensibly accurate estimation of robot’s pose in extremely dynamic, cluttered indoors environments. Design/methodology/approach The parameter vector of the statistical model is in the form of positions of easily detectable artificial landmarks in omnidirectional images. First, using probabilistic principal component analysis, the most likely set of parameters of the environmental model are extracted from the sensor data set consisting of missing values. Next, we use these parameters to approximate a probability density function, using support vector regression that is able to calculate the robot’s pose vector in each state of the Markov localization. At the end, using this density function, a good approximation of conditional density associated with the observation model is made which leads to a sensibly accurate estimation of robot’s pose in extremely dynamic, cluttered indoors environment. Findings The authors validate their method in an indoor office environment with 34 unique artificial landmarks. Further, they show that the accuracy remains high, even when they significantly increase the dynamics of the environment. They also show that compared to those appearance-based localization methods that rely on image pixels, the proposed localization strategy is superior in terms of accuracy and speed of convergence to a global minima. Originality/value By using easily detectable, and rotation, scale invariant artificial landmarks and the maximum statistical power which is provided through the concept of missing data, the authors have succeeded in determining precise pose updates without requiring too many computational resources to analyze the omnidirectional images. In addition, the proposed approach significantly reduces the risk of getting stuck in a local minimum by eliminating the possibility of having similar states.