Automated Terrain Classification with a Bayesian Hyperparameter Optimized Deep Supervised Autoencoder Model

Abstract

Terrain classification according to specific terrain attributes, has become increasingly important in certain decision-making scenarios. Automated robots are often utilized to traverse a specific surface to collect data that can be used in classification models to identify a specific terrain. In this study, a supervised autoencoder model (i.e., an autoencoder combined with a supervised learner such as a multilayer perceptron) is proposed to perform the classification of different terrains. Furthermore, a Bayes hyperparameter optimization approach is employed to determine optimum hyperparameter values. The dataset used for model building and training was obtained by driving a Lego Mindstorm EV3 mobile robot, fitted with a Raspberry Pi computer and a Sense HAT inertial measurement unit over six different terrain surfaces, i.e., asphalt, dirt, epoxy, grass, paving, and stone surfaces. The final dataset contains 281 232 data points which were used for model building. The results of the proposed supervised autoencoder were compared and contextualized with three other models, i.e., an SVM model, a logistic regression model, and an XGBoost model. Results indicate that it is not only feasible but also desirable to consider the use of a supervised autoencoder model when there is a need for terrain classifications.