In the application of human activity recognition (HAR) based on channel state information (CSI), due to the high dynamic characteristics of wireless channel to different environments, the features of human activity samples in different locations are different. In addition, the existing CSI-based HAR approaches limit the extraction of activity features to the Euclidean space and ignores the rich relational information between samples, categories and locations, which result in insufficient generalization performance for location-independent HAR. To address this challenge, this paper proposes a CSI-based location-independent HAR system CSI-MTGN. The system represents the classification task under each training sample collection location (TSCL) as a task, which is composed of three interactive parts: sample hidden representation, activity features extraction based on hierarchical graph neural network (HGNN) and information exchange based on multi-task learning. The proposed system improves the sample hidden representation, which is benefit for activity feature extraction and classification. The HGNN is designed to express various relationship information between samples, categories and locations in the form of graph structure, and the classification task under each TSCL is constructed through data augmentation, so as to improve the knowledge understanding and inference capabilities of the recognition model. The multi-task learning is used to achieve implicit data augmentation by sharing parameters among tasks through soft parameter sharing, and improves the generalization performance of the system. To validate the performance of the proposed system, experiments were conducted in a hall and a conference room, where samples of 10 categories of activities under 7 TSCLs were used for training the system, and the HAR accuracy rates at any locations were 94.1% and 93.3%, respectively.

Human activity recognition (HAR) based on channel state information (CSI) plays an increasingly important role in the research of human–computer interaction. Many CSI HAR models based on traditional machine learning methods and deep learning methods have encountered two challenges. A lot of CSI activity data is needed to train the HAR models, which is time consuming. When the indoor environment or scene changes, the recognition accuracy of the model drops significantly, so it is necessary to recollect data to train the model. The existing few-shot learning-based method can solve the above problems to some extent, but when there are more kinds of new activities or fewer shots, the recognition accuracy will decrease significantly. In this article, considering the relationship between various activity data, a graph-based few-shot learning method with dual attention mechanism (CSI-GDAM) is proposed to perform CSI-based HAR. The model uses a feature extraction layer, including the convolutional block attention module (CBAM), to extract activity-related information in CSI data. The difference and inner product of the feature vector of the CSI activity samples are used to realize the graph convolutional network with a graph attention mechanism. The experiments proved that under the learning task of recognizing new activities in the new environment, the recognition accuracy rates reached 99.74% and 98.42% in the 5-way 5-shot and 5-way 1-shot cases, respectively. The proposed method is also compared with other few-shot learning and transfer learning methods.

Activity recognition based on channel state information (CSI) plays an increasingly important role in human computer interaction. However most CSI activity recognition systems need to re-collect a large amount of samples and retrain model when they are used in new environments or recognize new types of activities, which greatly reduces the practicality of CSI activity recognition. To address this problem we design an adaptable CSI activity recognition system based on meta-learning, which only needs to fine-tune model with very little train effort when it is used in new environments or recognize new types of activities. Specifically, we first use meta-learning algorithm to get the pre-trained model that adapts to task distribution, when the environment or activity category changes, our system doesn't need to retrain the model and has maximal performance after updates the pre-trained model through one or more gradient steps computed with a small amount of samples from new activities. To prevent the loss of CSI time information after feature extraction with multi-layer CNN, we add time encoding on CSI data as the input of CNN neural network. Considering that CSI data may be labeled incorrectly during labeling process, we improve categorical cross entropy loss(CCE) to enhance the system's robustness to these mislabeled data. We test our system on the gesture dataset and the body activity dataset, and the experimental results show that our system achieves average accuracy of 72 percent with one sample of each new activity and 89.6 percent with five samples of each new activity.