Efficient Skeleton-Based Human Action Recognition for IoT Edge Devices Using Lightweight LSTM Networks

Human Action Recognition (HAR) is becoming increasingly important in areas such as intelligent surveillance, healthcare monitoring, and facilitating human-computer interaction. However, most current techniques use RGB-based deep learning, which is computationally intensive and cannot run on low-resource edge devices. This study presents a lightweight, skeleton-based Human Action Recognition (HAR) framework. The framework utilizes MediaPipe pose estimation, spatiotemporal normalization, and LSTM-based temporal modeling to efficiently recognize kinetically distinct human movements on resource-limited Internet of Things (IoT) edge devices. It combines MediaPipe pose estimation with Long Short-Term Memory (LSTM) networks to model temporal motion efficiently in resource-constrained IoT edge computing environments. Twelve kinetically unique actions from the UCF50 dataset were rendered into skeletal representations by temporally interpolating and spatially normalizing the action sequences. The HAR framework achieved 85.14% classification accuracy, indicating that sufficient motion-related information can be inferred from skeletal topology, allowing for reliable recognition at a reduced computational cost. The results demonstrate a practical balance between efficiency and performance, positioning the framework as an effective tool for edge-oriented HAR applications. From an IoT perspective, the framework has the potential to support IoT-enabled smart surveillance, assisted living, and intelligent monitoring systems by making it suitable for connected cameras and edge devices for local action recognition, even with limited computing power and communication bandwidth. This makes it a promising candidate for deployment across diverse real-world environments while preserving privacy through skeleton-based representations.