Head Motion Classification for Single-Accelerometer Virtual Reality Hardware

Abstract

Head motions classification applied to virtual reality (VR) systems is still an open problem without a leading pattern recognition solution. In contrary to typical motion capture pattern recognition problem in this case we use only single inertial measurement unit (IMU) sensor. Head motions that we want to recognize in VR systems might be both natural head motions like nodding or shaking head (they might be used while interacting with VR avatars) and also elements of head-based navigation system or interface. The second type of actions is more challenging because it might contains actions that generate motion trajectories that do not appear in real-life, though they have to be possible to execute only by using a head. In this paper we propose a trajectory-based motion features description that is utilized by dynamic time warping (DTW) classificator. The training of the classificator requires using modified dynamic time warping barycenter averaging (DBA) heuristic algorithm which utilizes quaternions to represents rotations. The proposed pattern recognition system together with its evaluation on the set of head motions acquired by VR system is our original contribution. We have evaluated our method on dataset consisted of 8 types of motions performed by two persons (there are 160 motions samples). In leave-one-out evaluation we have obtained very good results: only 10% of one and 15% of another action has been incorrectly classified, while remaining 6 actions classes had been 100% correctly classified. Both dataset and implementation of proposed method can be downloaded, due to this our experiment can be reproduced.