Multi-user Pose Prediction Based on GAT-GRU Framework in Mobile Augmented Reality Systems

Mobile Augmented Reality (MAR) constitutes a transformative platform for multi-user collaborative applications. Nevertheless, its efficacy in multi-user contexts is significantly limited by the stringent requirements for maintaining high interaction consistency. Failure to meet these requirements often results in visual desynchronization and interaction breakdowns, which critically undermine user immersion and the overall collaborative experience. Moreover, conventional pose localization and prediction techniques generally depend on substantial computational resources, posing a fundamental conflict with the stringent low-latency and low-energy consumption demands characteristic of mobile AR applications. This challenge is further exacerbated by the high mobility and dynamic behavior of users within multi-user MAR systems, which can increase latency and consequently degrade the user experience and Quality of Experience (QoE). To address these challenges, we introduce a novel framework that leverages spatiotemporal modeling for pose prediction, integrating a Graph Attention Network with a Gated Recurrent Unit (GAT-GRU). This model is engineered to predict user poses by jointly capturing the complex spatiotemporal dynamics inherent in multi-user AR interactions. This predictive capability underpins a proactive and adaptive resource allocation strategy, enabling the system to anticipate and fulfill future resource demands rather than merely responding to current conditions. The proposed GAT-GRU model exhibits superior efficacy in pose prediction tasks, markedly surpassing baseline models such as GCN-GRU, Social-GRU, and St-gat across principal evaluation metrics. Specifically, for position prediction, the model attains a mean absolute error (MAE) of 0.901 m and a root mean square error (RMSE) of 1.421 m, which are significantly lower than those reported for the baseline approaches (e.g., MAE values of 3.739, 1.437, and 1.039 m). In terms of rotation angle prediction, the GAT-GRU model achieves an MAE of 0.250 rad and an RMSE of 0.337 rad, thereby outperforming the comparative models. These empirical findings provide strong evidence supporting the model’s effectiveness within multi-user augmented reality (MAR) environments.