Towards the Usage of Synthetic Data for Marker-Less Pose Estimation of Articulated Robots in RGB Images

Pose estimation is a necessity for many applications in robotics incorporating interaction between the robot and external camera-equipped devices, e.g. mobile robots or Augmented Reality devices. In the practice of monocular cameras, one mostly takes advantage of pose estimation through fiducial marker detection. We propose a novel approach for marker-less robot pose estimation through monocular cameras utilizing 2D keypoint detection and 3D keypoint determination through readings from the encoders and forward kinematics. In particular, 2D-3D point correspondences enable the pose estimation through solving the Perspective-n-Point problem for calibrated cameras. The method does not rely on any depth data or initializations. The robust 2D keypoint detection is implemented by modern Convolutional Neural Networks trained on different dataset configurations of real and synthetic data in order to quantitatively evaluate robustness, precision and data efficiency. We demonstrate that the method provides robust pose estimation for random joint poses and benchmark the performance of different (synthetic) dataset configurations. Furthermore, we compare the accuracies to marker pose estimation and give an outlook towards enhancements and realtime capability.