Rigid Body Head Tracking with Scattering: A Comparison of Linear, Gaussian Process and Deep Learning Models

Motivation: To explore the use of radio frequency sensors to quantify changes in head pose; to study the relationship between head position and scattering. Goal(s): To compare head motion tracking models to convert scattering matrices to motion parameters. Approach: A rapid series of EPI images were used as the ground truth. Linear, Gaussian Process and simple deep learning models were evaluated. Results: For 5 of the 6 datasets, all models were limited by the accuracy of the ground truth data. In the final dataset, we demonstrated Gaussian Processes and a simple network can provide better head tracking than linear models. Impact: We have shown that Gaussian Process and simple deep learning models can outperform the existing linear models for tracking head pose with scattering on a single subject. This could be valuable in the future for applications to motion correction.