Novelty detection is about recognising features that do not fit into the pattern of previous perceptions. It is an important survival trait for animals, and is also useful for robots. For example, a robot equipped with the ability to detect novelty can select which features of an environment to investigate and learn about. In this paper we enable a mobile robot to learn a model of an environment that the robot experiences through the images of a monochrome camera while exploring. Once the robot has learnt a model of this environment we move the robot to a new environment and ask it to detect those features of the new environment that do not fit into the model, i.e., the novel features. We describe a number of different algorithms for producing an input vector from the image that is suitable for presentation to the novelty filter, and demonstrate results using the approach that worked best.
展开▼
