VR Vertigo Level Classification Using a Multi-Dimensional Taylor Network Approach

Users often experience motion sickness symptoms after using Virtual Reality (VR) systems, which can jeopardize their health. If we can rapidly diagnose the vertigo levels of VR users and dynamically take anti-motion sickness measures based on the levels, such as adjusting the movement of objects in VR glasses, adding visual reference points, and changing image brightness, contrast, and refresh rate, we can quickly alleviate motion sickness symptoms. Therefore, rapid assessment of the VR vertigo levels becomes crucial. Deep learning methods can accurately diagnose vertigo levels, but due to the complex structure and high computational requirements of these deep models, they may not fully meet the need for speed. Additionally, these complex models may be challenging to implement in devices like VR glasses. Unlike deep neural networks, BP-MTN represents complex nonlinear functions as polynomial functions using a simple network structure based solely on addition and multiplication operations. This design significantly reduces model complexity. However, traditional MTN models are primarily used for prediction tasks and are not suitable for classification. To address this issue, this paper proposes a Backpropagation Multivariate Taylor Network (BP-MTN) classifier for diagnosing VR vertigo levels. Compared to the traditional MTN, the BP-MTN includes the following modifications: 1) adding fully connected layers to handle inconsistent input and output dimensions of the BP-MTN; 2) introducing a softmax layer after the traditional MTN's output layer to enable classification; 3) incorporating activation functions after each output node to enhance the model's ability for fitting nonlinearity. Experimental results demonstrate that the BP-MTN classifier achieves higher classification accuracy and faster speed compared to some deep learning models.