Terrain Optimized Nonholonomic Following of Vehicle Tracks

The estimation of the ground surface is essential for a mobile robot's safe traversal in an unknown environment. As a first step, terrain sections which exhibit similar features can be clustered together for establishing a broad structure of the underlying environment. In this work, we focus on an unsupervised learning approach to segment different terrain types according to the clustering of acquired vibration signals. We propose a Gaussian mixture model-based clustering approach taking the inherent temporal dependencies between consecutive measurements into account. Therefore, we combine the expectation maximization algorithm (EM) with the probabilistic framework of a Bayes filter. While the E-step of the EM algorithm determines the probability of each measurement to belong to a certain cluster, the Bayes technique then filters these probabilities over time for their later use in the M-step of the EM algorithm. In this context, time relates to the sequence in which the measurements occur during the robot traversal. The evaluation using data collected from our RWI ATRV-Jr robot shows that our approach generates stable models for a variety of robot driving speeds even in situations of high-frequency terrain changes.