Robotics and Autonomous Systems 58 (2010) 634–647 Contents lists available at ScienceDirect Robotics and Autonomous Systems journal homepage: www.elsevier.com/locate/robot Novel solutions for Global Urban Localization C.U. Dogruer a,∗ , A.B. Koku b , M. Dolen b a Mechanical Engineering Department, Hacettepe University, Ankara, Turkey b Mechanical Engineering Department, Middle East Technical University, Ankara, Turkey article info Article history: Received 6 January 2009 Received in revised form 29 November 2009 Accepted 3 December 2009 Available online 29 December 2009 Keywords: Outdoor localization Mixture of Gaussians Particle filter Viterbi algorithm Genetic algorithm Extended Kalman filter abstract In this study, novel solutions to Global Urban Localization problem is proposed and examined rigorously. Classical approaches including Particle Filter, mixture of Gaussians, as well as novel solutions like Viterbi Algorithm and differential evolution are evaluated. The contribution of this paper is twofold: The Viterbi algorithm is extended by exploiting the structure of the problem at hand that is the states are partially connected temporally. Differential evolution is modified by taking into account the covariance matrix of states. Thus states encoded in genes are only allowed to interact locally within the region described by covariance matrix. This prevents the differential evolution from getting trapped into false maxima in the early stages of optimization. Finally, it is demonstrated with extensive experiments that solution of Global Urban Localization problem is possible. © 2009 Elsevier B.V. All rights reserved. 1. Introduction Global Urban Localization (GUL), which is initially defined in [1,2], is a mobile robot localization problem which aims to localize a mobile robot on urbanized regions of the world. GUL implicitly refers to the iterative position estimation techniques (and their applications) for mobile robotic systems utilizing satellite images along with the consistent sensory data acquired by the robots through their courses in urbanized settings. In fact, the GUL, which can be regarded as a special case of GL, specifically deals with urbanized settings owing to the fact that such locales with bountiful geometric entities offer distinct sensory features enabling easy match to those of the images. In GUL, the satellite images serve as primary tools to provide a priori information about the environment. In this study, the so-called GUL problem is studied in detail. A number of solution methods are proposed and evaluated: Particle Filter, mixture of Gaussians, (constrained) Viterbi Algorithm, differential evolution method. As the baseline cases, Particle Filter and mixture of Gaussians, which are particularly tailored for the solution of GUL, are taken into consideration and are compared to those methods developed/modified in this study. Within the context of this work, the Viterbi Algorithm is modified by taking ∗ Corresponding author. Tel.: +90 312 2976208; fax: +90 312 2976206. E-mail addresses: cdogruer@hacettepe.edu.tr (C.U. Dogruer), kbugra@metu.edu.tr (A.B. Koku), dolen@metu.edu.tr (M. Dolen). into account the fact that the motion model imposes constraints on future states. It is shown that the states of hidden Markov model describing the navigation of a mobile robot are not fully connected in successive time steps. By exploiting this information, Viterbi Algorithm is accelerated. Next, the localization of a mobile robot is expressed as an optimization problem and differential evolution method is applied to solve this optimization problem e.g. localization of a mobile robot. However, this optimization problem cannot be solved in its naive form because of the complexity of the problem so an incremental approach is assumed and special attention is paid to propagate the local maxima to the next time step. Thus, the differential evolution method is altered by incorporating the island models which is inspired by the fact that covariance matrix of the state vector describes a specific elliptical region in which the states are most likely to interact. Hence the local maxima are found and maintained, until system converges to maximum, which is the true pose of mobile robot. It is critical to note that the laser range finder is used as the primary sensing sensor for such a large-scale outdoor localization problem and both motion and measurement models are presented in detail. The vehicle model e.g. motion model, measurement model i.e. how information extracted from satellite image can be used to localize a mobile robot in urban environment is studied. Thus, the objective of this paper is to evaluate the above- mentioned methods as potential solutions for the GUL problem. Evaluation is based on field tests which are to demonstrate the efficiency and unique side of these methods as a virtual GPS application. 0921-8890/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.robot.2009.12.001