Spatio-Temporal Modeling Methods for Point Cloud Video Based on Federated Learning

Point cloud videos, with their ability to capture dynamic scenes, have shown significant value in fields such as autonomous driving, robotic vision, and virtual reality. Current point cloud video modeling methods often require large amounts of data to be centralized on a server. However, increasingly strict data privacy regulations now prohibit such centralized storage, limiting the development of point cloud video technology. Existing federated learning models typically use fixed aggregation parameters, making them less adaptable to the recognition needs of point cloud videos. This paper proposes a federated learning-based point cloud video recognition system, FedPST, which integrates game theory concepts to design a client controller and a server controller, enabling dynamic generation of aggregation parameters for flexible fusion of point cloud video models. Through experiments on multiple datasets, we validate the effectiveness of FedPST in 3D human action recognition tasks. The main contributions of this paper include: proposing a federated learning-based spatiotemporal modeling system for point cloud videos, introducing a game theory-based controller mechanism to address statistical heterogeneity, and experimentally verifying the performance improvements of this system. Results show that FedPST significantly improves the accuracy and efficiency of point cloud video recognition while preserving data privacy.