Hand Pose Estimation from Depth Data

A key challenge in modern robotics and biomedical engineering is to design artificial hands able to reproduce human abilities [2]. The difficulty to handle human-like manipulation problems is mainly due to the high number of Degrees of Freedom (DOFs) concentrated in a small volume. As a consequence, the control of robotic grasp and manipulation is an interesting challenges for engineers and scientists in the fields of robotics and machine learning. A possible solution consists in learning manipulation tasks from human observation. The fist step to apply learning methods to control anthropomorphic hands consists in tracking human palm and the fingertips (contact points) from a camera sensor. The objective of this Practical Work is to develop an algorithm able to robustly estimate the poses of palm and fingertips using only depth data. To this end, a deep learning technique [1] will be used.

Tasks:
• Literature review on human hand tracking with depth data.
• Developing an algorithm to estimate the palm and fingertips pose.
• Evaluation on synthetic and real datasets.

Supervision: Christian Rupprecht, Federico Tombari
This project is in collaboration with Pietro Falco and Matteo Saveriano from the chair of Dynamic Human-Robot Interaction.

For further information please contact: Christian Rupprecht

Bibliography:
[1] Vasileios Belagiannis, Christian Rupprecht, Gustavo Carneiro, and Nassir Navab. Robust optimiza- tion for deep regression. In International Conference on Computer Vision (ICCV), 2015.
[2] Antonio Bicchi. Hands for dexterous manipulation and robust grasping: A difficult road toward simplicity. Robotics and Automation, IEEE Transactions on, 16(6):652–662, 2000.