Z-TRANSCGAN: A Z-SCORE NORMALIZED TRANSFORMER-BASED CONDITIONAL GAN FOR GAIT DATA AUGMENTATION

Abstract. Gait analysis is essential for disease diagnosis, identity verification, and rehabilitation assessment. However, the effectiveness of deep learning approaches in gait analysis is hindered by the difficulty of collecting large-scale, high-quality gait time-series data, which is costly, labor-intensive, and subject to strict privacy regulations. Experimental findings on the HOA dataset indicate that Z-TransCGAN surpasses TTS-CGAN, achieving a 1.42% increase in average classification accuracy (ACC) and a 0.85% increase in the area under the curve (AUC). These results validate the efficacy of Z-TransCGAN as a data augmentation strategy for gait analysis, improving both synthetic data generation and downstream classification performance. Keywords: Deep learning; Conditional Generative Adversarial Networks (CGANs); Gait analysis; Data augmentation; Transformer.