A Parallelized Framework for Human Action Recognition and Prediction Based on Graph Neural Networks

With the rapid development of intelligent technology, human action recognition and prediction have important applications in fields such as video surveillance, security, human-computer interaction. However, current approaches to human action recognition and prediction often involve separate training processes, which can lead to error propagation from the recognition module to the prediction module, increasing prediction errors. To address this, this paper proposes a parallelized framework based on graph neural networks (PRP), consisting of a backbone network, action recognition head, and action prediction head. First, features are extracted through graph convolutional layers in the backbone network; then, the action recognition head outputs the action category <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\hat{y}$</tex>; finally, the action prediction head uses <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\hat{y}$</tex> as prior information to predict future action sequences. Joint optimization enhances the model's learning capability and reduces error accumulation. Experimental results demonstrate that the PRP model significantly outperforms comparison algorithms and holds broader application prospects.