SOFTWARE FOR THE LEARNING OF MONTE CARLO LOCALIZATION STRATEGIES USING OMNIDIRECTIONAL IMAGES

The computation of the position of an autonomous agent in a given environment is an essential problem in mobile robotics, since the pose is needed for a precise navigation. We present a tool that deals with the problem of mobile robot localization using omnidirectional images and an Appearance-Based Monte Carlo approach. In this method, the scenes are stored without any feature extraction, and the localization is carried out applying a Monte Carlo Localization method. To represent the appearance of each omnidirectional image with invariance to the rotation, we have used a single Fourier descriptor named Fourier Signature. We can divide the approach in two phases:
A.Map building. The mobile robot crosses the environment to build the map, and takes some images from several points of view. Due the high dimensional data of the images, to extract the most relevant information, a compression phase is needed. We have used the Fourier Signature with this aim. After this phase, the map consists of a data vector for every location.
B.Monte Carlo Localization. To carry out a task in the environment, the robot has to compute its localization in the map. To do it, the robot performed a trajectory inside the map, gathering omnidirectional images. When the robot acquires a new image, compresses it and uses it with the odometry information to actualize the Monte Carlo Algorithm. As a result of the algorithm, the localization of the robot is obtained.

The software tool we present has been designed to be used in a mobile robotics and computer vision subject. This tool will provide a simple and intuitive mechanism for students to understand the appearance-based technique in robotics localization through a friendly user interface, with the next features:

•Several databases with grey-scale omnidirectional images of an environment are included. These images must be used to build the map when the student begins the experiment. To carry out the localization process, some test images are included. These images have been captured in a real environment, following different path to test the validity of the Monte Carlo Algorithm.
•The students can select different values of error variance in the motion data to simulate the odometry information.
•The student can decide the amount of information he wants to retain from each image.
•To test the performance of the Monte Carlo algorithm, the student can select the particle number and the number of associations used. Also, the student can select the type of weight used to update the position of the particles.
•This tool is completely interactive. As the robot moves through the map, the student can see the result of the Monte Carlo algorithm (the particles position) and two graphs that show the error in the location and dispersion of particles. Also, the student can see the position of the robot according to odometry.

This software tool has demonstrated to be very useful for the students to understand the appearance-based Monte Carlo Localization. The students understand why it is necessary to compress the information when an appearance-based method is applied. They learn the performance of the Monte Carlo Algorithm, the influence of the type of particles weight in the outcome of the algorithm and the importance of a correct data association. Once the practical sessions have been completed, the students are capable to apply and develop probabilistic algorithms to realize a localization of the robot using this approach.