6-DOF Object Tracking
A system for automated data collection and off-line pose estimation for predetermined objects for robotic grasping testbed
This project was part of a few months I spent at the Collaborative Robotics and Intelligent Systems Institute of Oregon State University. I was part of an NSF Research Experience for Undergrads, and my goal was to develop an automated system for tracking an object.
The larger goal was to build an entire testbed which researchers could remotely access and test robotic grasping algorithms on. My part of the project would allow for quantitative statistics on the grasping experiments.
To approach the problem I developed a Python library that was capable of interfacing with Intel RGB-D cameras and recording several image streams in parallel. To then perform object tracking several state-of-the-art pose estimation techniques were implemented and compared. The comparison showed improved reliability and performance under occlusion for a Latent-Class Hough Forest framework developed at the Imperial College London.
During a grasping experiment the data capture was set to be automated by the RGB-D cameras and the images were then post-processed by the chosen framework.