Real Time Noise Reduction to Identify Motion Parameters in AR Maintenance Scenario.

Augmented reality is a field which improves user experience of the real environment by providing some relevant additional data. Understanding what happens in the workspace of the AR system in a maintenance context and for checking compliance of workers actions according to the expected ones are major challenges in this field. Usually, proposed approaches in literature are user centred and consist to gestures classification techniques. We opt to object centred methods. Indeed, when models are well configured, they provide information about motion between parts implied in assembly tasks which could be compared to predefined motion constraints. However, in real conditions, extracted motion curves are very noisy, may be difficult to exploit and may induce some AR systems misinterpretations. In this paper, we propose a method to reduce noise in real time in these curves based on Support Vector Machines confidence scores. The goal is to appropriately weaken false values and correctly straighten rotation axis according to the confidence we could have on pose estimation. Preliminary results are promising but the method still needs some improvements.