The Markerless Motion Capture System takes markerless video inputs from a synchronised and calibrated multi-camera system and produces the required virtual marker 3D positions as if the markers were on the subject.
Current state-of-the-art, data-driven keypoint localisation methods have limitations in accuracy because they only rely on datasets from manual annotations of joint centre locations. Instead, the Markerless Motion Capture System uses directly measured marker positions to provide more accurate keypoint localisation.
This idea is currently being tested with a dataset that contains videos from synchronised and calibrated cameras together with 3D marker positions from a standard marker-based motion capture system.
- A multi-camera video recording system with calibration workflow must be engineered and built to work together with marker-based motion capture system to ensure the precision of the ground-truth projection in both space and time.
- A group of subjects will perform a variation of movements individually captured by the system.
- A convolution neural network will be a core component in the learning architecture for the learning of marker localisation.
If successful, this project will remove time-consuming elements of the motion capture such as placing markers or wearing obtrusive equipment and will make it more practical for medical-related uses.
If you are interested to collaborate with us, please contact us here.