LAR: A Liteweight Paradigm for Human Action Recognition

As the application of action recognition technology in intelligent monitoring, healthcare, and other fields continues to expand, there is a growing demand for efficient, low-power action recognition models. Traditional RGB image-based action recognition paradigms often rely on complex Transformer architectures to directly detect the type of action, which not only has a high computational cost but is also difficult to implement on resource-constrained devices. To overcome these limitations, this study proposes a lightweight action recognition paradigm (LAR) aimed at achieving efficient and accurate action detection while reducing the demand for hardware resources. Our approach first utilizes the RTMPose model for human key point detection, which can accurately and quickly extract skeletal information from images. Subsequently, the detected key points are input into the PoseC3D model, which is specifically designed for action classification based on skeletal point data. Compared to existing technologies, our lightweight paradigm has significant advantages: it greatly reduces the number of parameters, speeds up inference, and significantly reduces the dependency on computing resources while maintaining high accuracy.