ROUGH-TERRAIN MOBILE ROBOT LOCALIZATION USING STEREOVISION Annalisa Milella Institute of Intelligent Systems for Automation National Research Council Via Amendola 122 D/O, 70126 Bari, Italy milella@ba.issia.cnr.it Giulio Reina Department of Innovation Engineering University of Salento Via Monteroni, 73100 Lecce, Italy giulio.reina@unile.it ABSTRACT Mobile robots are increasingly being used in high-risk rough terrain situations, such as reconnaissance, planetary exploration, safety and rescue applications. Conventional localization algorithms are not well suited to rough terrain, since sensor drift and the dynamic effects occurring at wheel- terrain interface, such as slipping and sinkage, largely compromise their accuracy. In this paper, we follow a novel approach for 6-DoF ego-motion estimation, using stereovision. It integrates image intensity information and 3D stereo data within an Iterative Closest Point (ICP) scheme. Neither a-priori knowledge of the motion and the terrain properties nor inputs from other sensors are required, while the only assumption is that the scene always contains visually distinctive features, which can be tracked over subsequent stereo pairs. This generates what is usually referred to as visual odometry. The paper details the various steps of the algorithm and presents the results of experimental tests performed with an all-terrain rover, proving the method to be effective and robust. 1. INTRODUCTION In order for a mobile robot to navigate autonomously over long distances on uneven surfaces, a method for accurately tracking the pose of the robot is primarily needed. Dead reckoning, based on data coming from wheel encoders, is a widely used localization method. This technique is easy to implement, and allows good short-term accuracy and very high sampling rate. However, dead reckoning systems are not well suited to long-range navigation and rough terrains, since they generally do not consider the physical characteristics of the vehicle and of the terrain it is traversing. Moreover, wheel slippage, sinkage, and sensor drift may cause errors that accumulate without bound over time unless an additional absolute localization system is employed for sporadic robot position updates [1, 2]. In this work, we follow a different approach, called visual odometry or ego-motion [3]. The basic idea of visual odometry is that of estimating the motion of the robot by tracking features of the environment detected with an on-board camera. Similarly to conventional dead reckoning, this technique can lead to error accumulation. However, since video sensors are exteroceptive devices, that is, they acquire information from the robot’s environment, visual odometry is not affected by wheel slippage and sinkage. Moreover, it has been demonstrated that vision allows better results for most sensor combinations [4, 5]. Several visual odometry methods have been proposed in the last decades, using either single cameras [5, 6, 7] or stereo vision [4, 7, 8, 9], which mainly differ depending on the feature tracking method and the transformation applied for estimating the camera motion. For instance, in [4], odometry provides an estimation of the approximate robot motion that allows a search area to be selected for improved feature tracking, and a maximum- likelihood formulation is employed for motion computation. In [5], the visual module uses a variation of Benedetti and Perona’s algorithm for feature detection, and correlation for feature tracking. Robustness is improved by integrating visual data with IMU using a Kalman filter. Finally, in [7], robust visual motion estimation is achieved using preemptive RANSAC [10], followed by iterative refinement. In this paper, an algorithm for 6-DoF ego-motion estimation is proposed, which incorporates image intensity information and 3D stereo data in the well-known Iterative Closest Point scheme. ICP was originally introduced by Besl [11], for the registration of digitized data from a rigid object with an idealized geometric model. Here, the potentialities of ICP are investigated for the case of visual odometry using stereovision. Specifically, two basic problems of ICP are addressed: the susceptibility to outliers, and the failure when dealing with large displacements. As an extension of these issues, another drawback of ICP is its inability to segment input data [11]. Typical solutions use odometry information for 1 Copyright © 2007 by ASME Proceedings of IMECE2007 2007 ASME International Mechanical Engineering Congress and Exposition November 11-15, 2007, Seattle, Washington, USA IMECE2007-41397