Human Action Recognition On Edge Devices: A Novel Light-Weight Model

Human action recognition (HAR) is an evolving technology with the potential to revolutionize how we understand human behavior, which finds applications across various domains such as elderly care, surveillance systems, and human-robot interaction. As HAR continues to advance, there's a growing interest in integrating it into Internet of Things (IoT) systems. To minimize response time between clients and servers, researchers have explored embedding models into edge devices, yet achieving optimal results remains a challenge. Balancing model size and performance is particularly problematic; while reducing parameters can limit model complexity, larger models often yield superior performance, posing challenges for implementation on memory-constrained edge devices. In this paper, we introduce a novel lightweight framework specifically designed to address these challenges. Through experimentation on a renowned benchmark dataset (JHMDB), our proposed approach demonstrates both superior performance and minimal model size.