Skeleton-Based Human Activities Fine-grained Recognition with RFID Technology
Meng Liu, Y J Chen, Hao Zheng, Lingbo Huang, Kun Zhao, Yilin Zhao · 2024
Human activity recognition(HAR) has received increasing attention and has been applied in multiple fields such as healthcare and human-computer interaction. Previous activity recognition methods have problems such as privacy leakage, strong intrusion on users, and coarse detection granularity. Therefore, we propose a privacy protection, easy-to-use, identifiability, lightweight, and fine-grained human activity recognition method based on bound-RFID technology. Firstly, we utilize the fundamental physical characteristics of radio frequency signals, such as doppler frequency (DF), received signal strength indicator (RSSI), and phase, which reflect human activity. Analyze the correlation between these data and human activities, establish a graphical relationship between data changes and activities, and generate human postures through posture time windows. Secondly, we integrate tags information to model human activities by establishing spatio-temporal skeleton graphs with temporal and spatial information. Finally, we model the spatio-temporal skeleton graph convolutional neural network to classify these graphs. As far as we know, this is the most refined HAR based on bound-RFID tags.