Research on Human Action Recognition Based on Edge Intelligence

Intelligent recognition algorithms deployed on edge devices offer strong real-time processing capabilities and high security for online video image analysis. However, real-time video image recognition remains challenging, particularly when deploying computationally intensive action recognition models on resource-constrained edge devices. This paper introduces a dual-network human action recognition method that leverages the strengths of both 2D and 3D convolutional neural networks (CNNs). Initially, a 2D CNN detects hu-man targets, followed by the generation of time-sequence data for these tar-gets using a matching method. Human action recognition is then achieved through a 3D CNN. This method addresses the challenges of extracting and modeling temporal action characteristics, enabling accurate human action localization and classification. Five distinct action recognition datasets were constructed to train the 3D CNN, which were subsequently deployed on an edge intelligent platform. Empirical testing of the action recognition system demonstrated a recognition speed of 10 frames per second, with accuracy exceeding 80% for certain action categories.