Image-Based Supervision of a Periodically Working Machine

Most industrial robots perform a periodically repeating
choreography. Our aim is to detect disturbances of
such a periodic process by a visual inspection system that can
be trained with a minimum of human effort and interaction.
We present a solution that monitors the robot with a time-of-
flight 3D camera. Our system can be trained using a few
unperturbed cycles of the periodic process. More specifically,
principal components are used to find a low-dimensional approximation
of each frame, and a One-Class Support Vector
Machine is used for one-class learning. We propose a novel
scheme for automatic parameter tuning, which exploits the
fact that successive images of the training class should be
close in feature space. We present exemplary results for a
miniature robot setup. The proposed strategy does not require
prior information on the dimensions of the machine or its maneuvering
range. The entire system is appearance-based and
hence does not need access to the robot's internal coordinates.