Learning the floor type for automated detection of dirt spots for robotic floor cleaning using gaussian mixture models

Abstract

While small floor cleaning robots rather cover area than detect actual dirt, larger floor cleaning robots in commercial settings need to actively detect and clean dirt spots. Floor types that have a single colour or simple texture could be tackled with an approach based on a fixed pattern. However, this restricts the use of the robots considerably. It terms of ease-of-use it is desirable to automatically adapt to a new floor type while still detecting dirt spots. We approach this problem as a one class classification problem and exploit the capability of the Gaussian Mixture Model (GMM) for learning the floor pattern. The advantage of the method is that it operates in an unsupervised way, which allows to adapt to new floor types while moving. An extensive evaluation shows that our method detects dirt spots on different floor types and that it outperforms state-of-the-art approaches especially for floor types with a high-frequency texture.