Cross-Domain Human Activity Recognition via Transfer Learning
Yong Zheng Zhou, Yilin Dong · 2024
In recent years, human activity recognition (HAR) based on wearable sensors has gained increasing attention, especially in the fields of medical health monitoring and exercise management. However, to achieve precise activity recognition, it is necessary to collect a sufficient amount of labeled data, which can be a time-consuming and labor-intensive task for data collectors. To address this issue, transfer learning utilizes labeled samples from a source domain to transfer knowledge to a target domain with few or no labeled samples. In this paper, we introduce the Transfer Component Analysis-based Long Short Term Memory (TCA-LSTM) network model, consisting of two core components. Firstly, the raw data is fed into the TCA module, resulting in a new feature representation where the differences between the source and target domains are minimized. Subsequently, this transformed feature representation is input into the LSTM module to achieve the final recognition of transfer behaviors. Experimental results on two open-source datasets demonstrate that the proposed TCA-LSTM outperforms other commonly used methods significantly.