Effective Classification of Head Motion Trajectories in Virtual Reality Using Time-Series Methods

In this paper, we present a method to classify head motion trajectories using recent time-series methods. The analysis of motion data with machine learning is a common technique to solve problems in Virtual Reality (VR), such as adaptive rendering or user behavioral modeling. Motion data are initially collected as time series, but they are usually transformed into tabular features compatible with traditional feature-based classifiers. Data mining research has proposed several time-series classifiers that can directly exploit the temporal relationship of the data without requiring manual feature extraction. Nevertheless, the effectiveness of these time-series methods still requires validation on real-life problem domains. Therefore, this paper demonstrates how a pipeline that combines a recent time-series classifier with two rotation space representations (quaternion and Euler) can successfully analyze head motion in VR applications. We test the proposed method on two public datasets containing head rotations, resulting in higher prediction accuracy than other feature-based and time-series classifiers. We also discuss some limitations, guidelines for future work, and concluding remarks.