SCFormer: A Vision Transformer with Split Channel in Sitting Posture Recognition

Prolonged maintenance of poor sitting posture can have detrimental effects on human health. Thus, maintaining a healthy sitting posture is crucial for individuals who spend long durations sitting. The recent Vision Transformer (ViT) models have shown promising results in various computer vision tasks. However, it faces challenges such as limited receptive field and excessive parameter quantity. To tackle these issues, we propose SCFormer. To begin with, we utilize the Regular Split Channel (RSC) module to partition the feature map along the channel dimension using specific rules. This enables the flow of spatial information within the channel dimension while severing positional information between adjacent channels, ultimately improving the model’s generalization. To extract local feature information and reduce computational complexity, we employ striped windows with parallel self-attention mechanisms over a subset of channels in the feature map. Lastly, we introduce Global Window Feedback (GWF), which exploits redundant information within the channel dimension through simple linear operations, enabling the extraction of inter-window global information and expanding the receptive field. By incorporating these design elements and employing a hierarchical structure, SCFormer demonstrates competitive performance in sitting posture recognition tasks. We achieve successful identification of 10 classes of sitting postures on our dataset, attaining an accuracy of 95%, surpassing current state-of-the-art models.