Human Activity Recognition Across Scenes and Categories Based on CSI

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.