From Temporal to Spatial: A Transformer-GAN Approach for Fall Prediction

The study aims to construct a fall detection system. We propose a Transformer-based Generative Adversarial Network (GAN) trained on human keypoint skeleton features. We arrange temporal information spatially, allowing the attention layers to learn movement details. Moreover, the objective of the GAN is to generate corresponding future motion sequences based on the current data. Consequently, the model can predict the next movement and determine whether an individual in the scene has fallen. Finally, the fall detection process employs distance metrics, including Chebyshev distance and Dynamic Time Warping (DTW), with a threshold determined through Bayesian decision theory. Experimental results demonstrate the model’s robust performance, achieving an F1-score of approximately 0.93 when applied to the evaluation set using the Chebyshev distance threshold. The framework effectively distinguishes fall and non-fall instances, showing particular strength in handling challenging scenarios like transitional movements.