Motion image recognition is a critical component of Internet of Things (IoT) applications, necessitating advanced processing techniques for spatiotemporal data. Conventional feedforward neural networks (FNNs) often fail to effectively capture temporal dependencies. In this work, we propose an indium gallium zinc oxide (IGZO) thin-film transistor (TFT) gated by a hafnium oxide (HfOx) dielectric layer, exhibiting voltage-modulated fading memory dynamics. The device exhibits transient current responses induced by oxygen vacancy migration, dynamically modulating channel conductance and enabling the transformation of 4-bit time-series sequences into 16 distinct states. This approach enhances the feature extraction process for motion history images by balancing the transient decay of individual frame contributions with the cumulative effect of the motion sequence. Systematic evaluation identifies an optimal pulse height of 2.5 V, achieving a motion direction classification accuracy of 93.9%. In contrast, simulations under non-volatile memory conditions exhibit static retention, leading to symmetric trajectories and significantly lower classification accuracy (49.6%). To further improve temporal data processing, we introduce the degree of state separation (DS) as a metric to quantify state distribution uniformity and identify optimal pulse conditions. This work advances the development of neuromorphic devices for efficient time-series data processing, providing valuable insights into the interplay between fading memory dynamics and neural network performance.
