Human behaviour prediction is important to enable
proper and safe collaboration between humans and robots
working within the same workspace. In an industrial environment
where not all the tasks can be automated, a human worker and
a robot can be performing tasks simultaneously in a common
workspace e.g. picking pieces from the same conveyor belt. In this
case the conventional approach of stopping the robot whenever
it is going to collide with the worker is not an optimal solution as
it decreases the actual cycle time of the robot and therefore the
productivity. In this paper a way of learning the intentions of the
human worker, so that the robot can still operate safely, is sought
based on 3D-sensor feedback, a skeleton model of the human and
a supervised classification algorithm. Partitioning Around Medoids
(PAM) is used to separate the recorded tracks into classes and
then SVM is trained with these data. Subsequently, real-time
recording is classified with the trained model to allow early
prediction of the action the operator is performing.