GAN-Enhanced Visual Transformer Approach for Thermal Sitting Posture Classification

Infrared thermal trace analysis captures residual heat patterns left by human bodies on seat surfaces, enabling post-event seating posture inference with inherent advantages of privacy protection and lightweight implementation. To further enhance generalization capability while preserving privacy-friendly characteristics, this study introduces a generative augmentation-driven approach for infrared thermal-based posture recognition. We collected 360 thermal images of 9 seated postures from 40 subjects. Using the FastGAN generative model, we expanded the dataset, screening to form over 8,106 high-quality synthetic images. A Vision transformer model was trained on this expanded dataset and compared with a model trained on the original data. Experimental results demonstrate that the jointly trained model using GAN-augmented and original datasets achieves a 30% improvement in classification accuracy, validating the feasibility of this approach for thermal seated posture recognition.