A Transformer-Based Approach to Human Posture Classification with 3D Skeleton Data

This study proposes a novel approach for human activity recognition using the Pose Sequence Transformer Neural Network (PoSeqTNet), trained using 3D skeleton input data. We focus on identifying the diverse postures involved in executing specific activities based on the spatiotemporal progression of an observation window. It contributes to the broader aim of recognising and predicting human activities through analysing human movements. We investigate the impact of random and sequential data inputs on postures classification performance. In total, 11 postures were evaluated. PoSeqTNet achieves good results in scenarios with smaller datasets, offering a noteworthy reduction in time and computational demands compared to classical Neural Networks, such as the clustering networks used in our other study. The results provide a highly advantageous balance between high performance, low computational requirement, and time efficiency, surpassing the capabilities of many state-of-the-art methods.