Vision-based localization for mobile robots

Abstract

Robust self-localization and repositioning strategies are essential capabilities for robots operating in highly dynamic environments. Environments are considered highly dynamic, if objects of interest move continuously and quickly, and if chances of hitting or getting hit by other robots are quite significant. An outstanding example for such an environment is provided by RoboCup.

Vision system designs for robots used in RoboCup are based on several approaches, aimed at fitting both the mechanical characteristics of the players and the strategies and operations that the different roles or playing situations may require. This paper reviews three approaches to vision-based self-localization used in the RoboCup middle-size league competition and describes the results they achieve in the robot soccer environment for which they have been designed.

Introduction

To perform complex navigation tasks and to coordinate its movement with other agents in indoor environments, an autonomous mobile robot needs knowledge about the world in which it is moving. This enables it to recognize the main landmarks and determine its position with respect to the environment. In fact, answering the question “Where am I?” is one of the most crucial and, often, critical tasks (both in terms of quality and time complexity) for an autonomous mobile robot [6], [12]. The self-localization problem is usually solved with a combination of relative and absolute position measurement techniques [7]. Odometry and inertial navigation (dead reckoning) systems are usually integrated with methods employing active beacons, natural or artificial landmarks as well as environment model matching, in order to compensate for wheel slippage and other positioning errors [13]. Natural and artificial landmarks are distinct features that a robot can recognize from its sensory input. In general, landmarks have a fixed well-known position within an absolute reference system, relative to which a robot can localize itself. The main task in self-localization is then to recognize such landmarks reliably by processing the sensory input to compute the robot’s position [4], [26], [30].

The literature concerning computer vision reports on a large number of natural and artificial landmark positioning systems integrated in autonomous robot architectures [3], [8]. Several authors have tried to find a good correlation between what the robot perceives through a camera and its internal representation of the environment [21], [27]. In indoor environments, edges of doors or walls are commonly used as landmarks to perform this correlation, but often such patterns are too generic to allow an accurate localization in a more complex environment. Researchers have often focused on the choice of optimal shapes or patterns as landmarks so that robust localization can be obtained from a single [23], [24] or from multiple images [20]. Furthermore, considerable work has been devoted to the treatment of error in sensor observations [5]. Sometimes it can be useful to evaluate alternative techniques that can recognize landmarks which can be specially designed to carry both human- and robot-oriented visual information, though they might also be complemented by other natural signs which are often found in office and industrial environments [2], [28]. Often these methods do not yield high localization accuracy, but there is a number of practical situations, e.g., docking and navigating in small openings, where coarser robot positioning accuracy suffices.

The RoboCup environment provides researchers with a common environment in which it is possible to test and compare different visual self-localization methods reliably. Self-localization in RoboCup is very important to achieve coordination between players, both to defend the team’s penalty area and to make goal-oriented decisions [19]. Although the world of RoboCup is simpler than many real-life indoor environments, it is a very dynamic world in which the robot has to react quickly to the changes caused by the continuous movement of the robots and the ball. Most robots rely on odometry for real-time self-localization. However, odometry fails immediately when a robot hits other robots. The appeal of the RoboCup challenge is actually due to its highly dynamic environment, in which robots usually obtain only partial views of the scene, since reference markers are often occluded by other robots.

To reposition themselves, robots in RoboCup use different kinds of sensors, such as cameras, laser range finders, sonars, and so on. While self-localization based on laser range finders provides a very accurate and robust estimation of robot position, visual sensors are more flexible; the number and configuration of cameras can be chosen according to which other tasks are performed by the vision system and to the role that the robot has within the team. Most distance sensors require that the field of play be surrounded by walls. This limits their future use, when RoboCup eventually moves towards a more natural, wall-free environment.

In RoboCup, according to the current rules, the appearance of relevant objects is well known and it is therefore possible to provide robots with some geometric models of the field or of its subparts. Visual self-localization is typically performed by matching and comparing the robot’s view of the world with a model. This comparison can often be used to refine and compensate for the errors introduced by odometric data, but in some cases a robot must be able to localize itself without any help from sensory systems other than that of vision.

Several kinds of vision systems have been applied in RoboCup. While most robots use traditional single-camera vision systems, it is becoming quite common to adopt more complex solutions. For example, the problem of having only a limited field of view can be solved using a camera pointing upwards to a concave mirror to obtain a global vision system. Sometimes, different vision system designs and localization strategies are used for robots playing on the same team, which allows for more specialized robots to be built according to their roles but makes sharing data among robots more difficult.

This paper reviews three approaches to visual self-localization used in the RoboCup middle-size (F-2000) competition, describing and comparing the results they achieve in the RoboCup test-bed environment. The paper is organized as follows: a self-localization method based on an omnidirectional vision system is illustrated in Section 2, a self-localization based only on the comparison between a pair of images from a binocular vision system is proposed in Section 3, and a self-localization based on the Monte Carlo localization technique is described in Section 4. Section 5 reports some results of the above described techniques. Finally some conclusions from which possible future work can stem are drawn in Section 6.