Abstract: Action recognition is vital for various real-world applications, yet its implementation on embedded systems or edge devices faces challenges due to limited computing and memory resources. Our goal is to facilitate lightweight action recognition on embedded systems by utilizing skeleton-based techniques, which naturally require less computing and memory resources. To achieve this, we propose innovative methodologies and optimizations tailored for embedded deployment, including post-training quantization, optimized model architectures, and efficient resource utilization. By enabling real-time and lightweight action recognition on resource-constrained embedded systems, our research opens up new possibilities for applications in areas like autonomous surveillance, driving, and indoor safety
monitoring.
Loading