Page - 16 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Visual Localization System for Agricultural Vehicles in GPS-Obstructed Environments* Stefan Gadringer1, Christoph Sto¨ger1 and Florian Hammer2 Abstract—Accurate outdoor localization and orientation de- termination using the Global Positioning System (GPS) usually works well as long as the GPS antenna receives signals from a sufficient number of satellites. Especially in agricultural appli- cations, the respective lines of sight are frequently obstructed due to the presence of trees. In this paper, we investigate the applicability of an alternative method for position and orientation estimation that is based on a stereo-camera system and Visual Odometry (VO). We have experimentally validated our approach in a logging road scenario. Based on the results of the position and orientation estimation, we discuss challenges of VO in such a non-trivial environment. I. INTRODUCTION Localization of a vehicle is a very important task and hence a research topic for decades. In general, localization is possible with sensors like GPS, rotary encoder, IMU (Inertial Measurement Unit), laser scanner or a camera. Of course, there exist even more sensors and each one has its own pros and cons in terms of accuracy, drift, price, etc. The area of application highly depends on these properties. In this paper, we focus on outdoor localization in natural terrain. This is an important topic for precision farming [4], for example. Hereby, the question is always the same: Which sensors are suitable for the application? Asdiscussed in [25], aGPSantennaalwaysneeds intervis- ibility to several satellites to guarantee an accurate position estimation. This is sometimes impossible in areas like in a forest where trees occlude the satellites. The usage of wheel odometry via rotary encoders is not suitable as well due to problems with inaccuracies of the wheel geometry and slipping situations. In comparison, an IMU allows a good estimation of the orientation but not for the position because the double integration of the acceleration results in a high drift over time. A laser scanner has a very high position accuracy on the one hand but it is very expensive and not so well proofed for high vibrations on the other hand. Thus, just the camera remains of the sensors mentioned above. This sensor is relatively cheap but a position and orientation estimation via VO is normally linked with high computing demand and continuous growth of the drift per number of used images. Furthermore, overexposed images and other problems like branches that occlude cameras need a robust *Parts of this work have been supported by the Austrian COMET-K2 programme of the Linz Center of Mechatronics (LCM), and was funded by the Austrian federal government, and the federal state of Upper Austria. 1Stefan Gadringer and Christoph Sto¨ger are with the Institute of Robotics, Johannes Kepler University, 4040 Linz, Austria {stefan.gadringer,christoph.stoeger}@jku.at 2Florian Hammer is with the Linz Center of Mechatronics GmbH, 4040 Linz, Austriaflorian.hammer@lcm.at implementation of a VO to be able to get a valid pose estimation. However, this paper shall show the applicability of Visual Odometry to estimate position and orientation in different wooden environments with ambiguous natural structures. This paper is structured as follows. Section II gives an overview of related work. Visual Odometry and all its com- ponents are explained in Section III. Finally, the experiments are shown in Section IV. Last but not least, Section V contains the conclusion as well as some remarks about future work. II. RELATED WORK Visual Odometry (VO) is the incremental estimation of the pose (position & orientation) via examination of the changes on images due to motion induction [24]. The research on VO already started in the early 1980s and one if its advantage is that no prior knowledge about the environment isnecessary.Agoodexample is the implementationofCheng et al. [6], [21], which was used in the rover of the NASA Mars exploration program. Since then VO was continuously under research, which means that the literature about Visual Odometry is huge. Therefore, this section just contains an overview about relevant literature of VO for the localization of a vehicle in an outdoor environment. Nister et al. [22] proposed one of the first real-time VO which was capable of a robust pose estimation over a long track. They use a stereo-camera system and detect Harris corner features [15] in the images. 3D points are estimated through triangulation of the corresponding features in a stereo pair. In a next step Nister et al. use these 3D points and the features of a following image to estimate the pose via a 3D-to-2D algorithm as described in [24]. RANSAC (Random Sample Consensus) [12] removes outliers in the motion estimation step. Regarding to Scaramuzza et al. [24], this VO procedure was a high improvement to previous implementations and is still used by many researcher. Comport et al. [7] use a similar procedure but estimate the motion using 2D-to-2D instead of 3D-to-2D feature correspondences. With reference to Scaramuzza et al. this results in a more accurate pose because triangulation is not needed. In [26], [17]or [27]bundleadjustment is applied to further reduce the drift of the Visual Odometry. Bundle adjustment optimizes the latest estimated poses using features over more than just two stereo pairs. Konolige et al. [17] show that this step reduces the final position error about a factor of two to five. 16