Image-Based Visual Servoing of Rotationally Invariant Objects Using a U-Net Prediction

Abstract

In this article an image-based visual servoing for the armature of electric motors is presented. For a calibrated monocular eye-in-hand camera system our goal is to move the camera to the desired position with respect to the armature. For this purpose we minimize the error between a corresponding feature vector and a measured feature vector. In this paper we derived various features from the output of a U-Net. The variety leads to the fact that we can decouple the features in the control process. The prediction of the U-Net is stabilized by strong augmentation, an armature model and an adaptive digital zoom. We can show that our U-Net control approach converges and is robust against noise and multiple objects.