A Low Cost Vision Based Localization System for Mobile Robots

Abstract

|Arti cial vision is one of the most versatile sensory systems. It can be used in many environments such as indoor, outdoor, space, and even in underwater contexts. Most of times, vision based localization requires complex algorithms and hardware resources when related to general environment features. However, using simple landmarks can reduce dramatically the cost and the complexity of the recognition system. In typical indoor environments, in particular in o±ces, ceiling lamps are all of the same type and are placed in a quite regular way. Moreover, they can be easily seen, as generally no obstacles can exist between them and the robot vision system. These peculiarities motivated a study on the possibility of implementing a very low cost localization procedure using a standard onboard webcam. Inexpensive hardware implies several problems: among the others, the need for a procedure that compensate for optical distortions, the poor quality of image, and the slow transfer rate. The paper describes the results of this study, with an emphasis on the implementation issues. Keywords|Localization, Vision, Mobile Robots. IN the last years there has been an increasing interest around localization systems for autonomous mobile robots. More and more often vision is added to the localization system of mobile robots in order to reduce the odometric errors due to the incremental nature of those sensors. Indeed, the use of natural or arti cial landmarks allows both simple resetting of odometric errors, so that sensors can restart each time with the correct absolute position information [1], and more complex sensor fusion algorithms that integrates all available information in a single estimation process [2]. Among the others, natural landmarks have the interesting features that avoid any change in the work environment. In indoor environments one can nd many di®erent reference points and marks that can be interpreted as natural landmarks, e.g. doors, corners, geometry of the °oor. In particular in [3] and [1] some of the ceiling characterizing elements are suggested as natural landmarks in order to avoid any occlusion problems. Unfortunately, visual based control schema requires, generally, very complex algorithm and dedicated hardware [4]. In this paper we have explored the possibility to realize a localization system for a mobile robot using very low cost standard hardware, i.e. a PC with a 40$ webcam. To this end we have mounted the webcam on the mobile robot focused to the ceiling and used the lamps as reference points. Moreover, we have used a suitable topological representation of the environments, i.e., we represent the environment by means of a graph where each node is a location of interest and the arcs capture the connectivity of the space. So the localization system uses the information about the position of each lamp to localize the robot on the graph and then inside the environment. Nevertheless, given the repetitive nature of the lamps, the system needs to label and track each lamp in the image sequence in order to avoid any wrong interpretation of the environment. This imposes an upper limit on the robot speed in order to guarantee that each lamp met by the robot on its path appears at least in one webcam image. Further, the presence of low cost hardware generates some drawbacks. The rst of all is due to the limited transfer rate that, in addition to the time consumed by the feature extraction algorithm, generates a considerable time delay in the loop. Also the poor quality of the image generated by the camera and the optical distortion introduced by its lens system should be taken into account especially when vision is used for precise positioning tasks. The presence of these limits have suggested the use of a two stage localization scheme: 2 While the robot is moving toward the target location, the algorithm uses a lower resolution set-up for the camera to maximize the frame rate throughput and relax the constraint on robot