Cloud-Edge Collaborative Human Activity Recognition System Design and Implementation

This paper proposes an edge computing architecture for Human Activity Recognition (HAR) based on Raspberry Pi and HaaS100, aimed at enhancing real-time performance and system robustness. The proposed approach employs Edge-End inference within the communication range of edge nodes, while relying on Cloud Edge inference otherwise, thereby establishing a comprehensive Cloud Edge End system that leverages the capabilities of a mobile communication 5G private network Wi-Fi. To achieve accurate activity recognition, three algorithms, namely Convolutional Neural Network (CNN), CNN-Long Short-Term Memory (CNN-LSTM), and Residual Network-based CNN (ResNet-based CNN), are implemented and their performances are rigorously compared using Actitracker datasets. The experimental results demonstrate that the ResNet-based CNN exhibits superior performance, achieving an impressive F1-score of 0.977. Additionally, to address the challenge of model size and resource-constrained edge devices, model compression is employed through channel-wise data collection, effectively reducing the parameter size from 220,000 to 160,000, with a negligible 0.5% decrease in accuracy. Finally, the compressed model is deployed on edge devices, and retraining is conducted using self-collected datasets to ensure practicality and adaptability in real-world scenarios. The proposed HAR edge computing architecture showcases significant advancements in real-time performance, system robustness, and resource optimization, making it a promising solution for activity recognition in edge computing environments.