Feature-Temporal Semi-Supervised Extreme Learning Machine for Robotic Terrain Classification

Robotic terrain classification refers to the ability that a field robot could identify the traversing terrains autonomously under as little human supervision as possible. Such a task could be achieved by semi-supervised learning which works in the premise of smoothness assumption in the feature space. However, we found that the feature smoothness assumption cannot be fully satisfied (i.e., there is no apparent low-density region in the feature space) in the robotic terrain classification, which motivates us to propose the feature-temporal semi-supervised extreme learning machine (FT-S2ELM). With introducing the feature-temporal similarity matrix, the accuracy of the classifier trained by semi-supervised learning increases significantly. Furthermore, considering the uncertainty in determining the smoothness degree (i.e., the free parameters of similarity matrix), we introduce an automatic approach to find the optimal graph Laplacian, thus increasing the safety. The proposed method is verified experimentally on the data gathered by a micro tracked robot.