USING AI FOR ACCURATE POSE ESTIMATION AND SERVOING
Pose estimation from camera data is challenging for a robot because traditionally camera data can be noisy, and without depth. A single camera image can't be relied upon for depth. Even if a stereo image is available, noisy data can lead to estimates that are 20cm-30cm off.
At the University of Freiburg, Thomas Brox and his team's work with a neural network yielded breakthrough results in tracking and pose estimation from monocular camera data.
The algorithm developed by the team is now also used by the robot's base camera ring on the robot body to achieve more accurate servoing as the robot positions and repositions itself around a plant to perform an even trim.
"Classical methods can fail in textureless things. Here we use AI to improve the robustness and also the accuracy."
Huizhong Zhou, PhD Researcher, Computer Vision and Geometry Group, University of Freiburg
Watch Huizhong talk about the University of Freiburg's research in the 39 second video clip below, taken from a short documentary on the TrimBot2020 project, Cutting Hedge Research, which you can watch by clicking here.
For more information contact:
Corin Campbell: Corin.Campbell@ed.ac.uk
Prof. Robert Fisher, Consortium Coordinator: rbf@inf.ed.ac.uk
