Non-rigid motion detection for motion tracking of the head

Abstract

Optical motion tracking systems are effective tools for measuring head motion during MRI and PET scans in order to correct for motion. Most systems rely on the attachment of fiducial markers which can slip or become decoupled from the head, causing erroneous motion estimates which can introduce further image artifacts. In this work, we investigated two methods of detecting non-rigid motion, both of which can be easily incorporated into a stereo-optical feature-based motion tracking system. The tracking system tracks detected features on small patches of the forehead. By monitoring these features, surface deformations on parts of the face that deform non-rigidly with respect to the rest of the head can be detected and potentially characterized. We investigated two methods of detecting non-rigid deformations: one involved measuring distances between detected landmarks and comparing these distances to previous frames; the other used a neural network to classify a group of landmarks as either `rigid' or `non-rigid'. A simulation tool was also developed to aid in the characterization of non-rigid motion and its effects. Landmark distance discrepancies were found to be correlated closely with pose measurement errors in the feature-based motion tracking system, suggesting it is a useful metric for detecting non-rigid motion. The trained neural network was able to classify a collection of landmarks as 'rigid' with 99.8 % accuracy and classified the `non-rigid' case with 93.3 % accuracy.