A Novel Approach to Recognising Physical Activity Intensity in Adults: Gravity-Based Acceleration Imaging with Vision Transformer

Research on physical activity intensity (PAI) recognition often relies on tabular data, limiting the capture of global patterns compared to image-based data, while traditional models struggle with long-sequence dependencies. This study introduces a novel method that uses a gravity-based acceleration algorithm to generate images of different intensity levels, followed by classification with a Vision Transformer (ViT) model, aiming to enhance model performance and experimental efficiency. The method uses the Euclidean Norm Minus One to calculate gravity-based acceleration per second. Gramian Angular Field images are generated based on adult intensity thresholds and are subsequently classified using the ViT model. The results showed the ViT model's high performance in recognising different PAIs. Specifically, it demonstrated that using gravity-based images for recognising PAI improves accuracy compared to METs-based images (etabolic equivalents), achieving an overall accuracy of 98.77% compared to 97.52%. Additionally, this novel approach maintains high accuracy across different time windows, with accuracy ranging from 98.59% to 99.05%, further indicating the model's stability and robust performance across varying temporal windows.