1 Comparison of Different Approaches to Vibration-based Terrain Classification Christian Weiss * Nikolas Fechner * Matthias Stark * Andreas Zell * * Department of Computer Science, University of T¨ ubingen, T¨ ubingen, Germany Abstract— There is a variety of different terrain types in outdoor environments, each posing different dangers to the robot and demanding a different driving style. In a previous paper, we presented a terrain classification method based on Support Vector Machines (SVM), which uses vibrations induced in the body of the robot to learn different terrain classes. However, in the previous paper, our experimental results were based on vibration data collected by a hand-pulled cart with relatively hard wheels. In this paper, we present experiments on data collected by our RWI ATRV-Jr outdoor robot. Additionally, we compare our SVM-based method to alternative classification methods. The comparison shows that our approach outperforms the other methods. Index Terms— Outdoor robotics, vibration-based terrain clas- sification I. I NTRODUCTION In outdoor environments, a mobile robot typically faces many different terrain types. Some of them are flat and not slippery, and therefore the robot can traverse them at relatively high speed. Other ground surfaces are loose, slippery or bumpy, and therefore dangerous. To prevent accidents, the robot has to traverse these regions slowly and carefully. These examples show that the ground surface itself is a possible hazard to the robot in outdoor environments. Such a hazard is called a non-geometric hazard [21]. The robot can only avoid accidents if it adapts its driving style to the current terrain type. One way to determine the terrain type is to directly estimate terrain parameters like cohesion or slippage from sensor mea- surements. Another way is to group the terrain into classes like asphalt, dirt or gravel, and to learn these classes from training examples. Once the robot has learned the different classes, it can classify new terrain data according to the learned model. The most common data used for terrain classification are data collected by laser scanners or cameras. Ladar-based methods often focus on segmenting the ground surface from vegetation or all kinds of obstacles (e.g. rocks) instead of estimating the type of the ground surface itself [18, 8, 19, 12]. Other methods divide the ground surface in navigable and non-navigable regions [22]. Vision-based methods usually use texture or color information [1, 4, 12]. Some research has also been done on using force-torque sensors and potentiometers to detect non-geometric hazards [9, 11]. Vibration-based terrain classification was first suggested by Iagnemma and Dubowsky [10]. The idea is to measure the vibration that is induced in the robot while it traverses the terrain. The vibration can be measured at the wheels, the axes or the body of the robot. Usually, accelerometers are used to measure the vibration perpendicular to the ground surface (z -acceleration). As different terrain types induce different vibration signals, one tries to learn characteristic vibration signals for each terrain type from training examples. The learned model is then used for classification of unknown data. The disadvantage of the method is that terrain can be classified only while the robot traverses it, but not beforehand. Advantages are, for example, the independence from illumina- tion conditions and the high reliability. Thus, vibration-based terrain classification can be used as a stand-alone classifier or in combination with other sensors. Brooks and Iagnemma examined vibration-based terrain classification for planetary rovers [2, 3]. They use Principal Component Analysis (PCA) to reduce the dimensionality of their data and Linear Discriminant Analysis (LDA) for classification. Sadhukhan and Moore presented an approach based on probabilistic neural networks (PNN) [14, 15]. In [20], we suggested an approach that uses Support Vector Machines (SVM) for classification. Stavens et al. presented an approach for vehicles driving up to 35 mph [17]. However, they focused on assessing the roughness of the terrain to adapt the velocity, and not on grouping the ground surface into classes. In our previous paper [20], we obtained our experimental results using data from a hand-pulled cart with relatively hard wheels. These wheels lead to relatively clear and strong vibration signals. In this paper, we present experimental results on data that we collected using our RWI ATRV-Jr outdoor robot. Its big, air-filled tires are likely to dampen the vibration signals. We also examine how different robot speeds influence the classification performance. Additionally, we implemented the terrain classification approach presented by Sadhukhan and Moore as well as the approach suggested by Brooks and Iagnemma, and compare both to our approach. These methods cover two of the four main groups of classification methods, namely kernel methods (our SVM-based approach) and neural networks (the PNN). From the third group, the methods based on Likelihood, we chose Na¨ ıve Bayes, which is a standard method from this group. The fourth group are decision trees, from which we examined the J4.8 algorithm. J4.8 is based on the well known C4.5 algorithm. Finally, we also tested a k-nearest-neighbor (kNN) classifier. The rest of this paper is organized as follows. Section II recapitulates our SVM-based terrain classification approach. Section III briefly describes the alternative classification meth- ods. Section IV presents our experimental results and finally, Section V concludes the paper and suggests future work.