Enhancing RFID based human activity recognition through spatiotemporal feature construction

This article proposes a new method that significantly improves the accuracy of action recognition based on RFID systems by introducing advanced spatiotemporal features and utilizing spatiotemporal graph convolutional networks (ST-GCN). Traditional RFID motion recognition models often struggle to capture the subtle spatial and temporal relationships between joint movements, resulting in limited recognition accuracy in complex activity scenarios. To overcome this issue, we propose a new feature construction method that utilizes new signal features derived from RFID signal data (including RSSI, Doppler frequency, and phase), such as velocity, acceleration, phase change rate, and RSSI change rate. These features can finely capture the dynamic changes in time and space, thereby improving the performance of the model in distinguishing complex actions. The experimental results show that the ST-GCN model trained on enhanced feature sets significantly improves the accuracy of action recognition tasks, verifying the effectiveness of our method. In addition, feature construction methods have played a key role in improving the identification ability of RFID signal data, especially in distinguishing similar actions in complex environments, demonstrating their broad potential for application in behavior analysis.