Introduction
Human Activity Recognition (HAR) plays a very important role in smart home, healthcare, fall detection, and VR/AR. Traditional HAR is achieved by computer vision technology or wearable sensors. Traditional HAR is achieved by computer vision technology or wearable sensors, where the former requires human objects must appear in the camera's Field of Vision (FoV) while the latter requires a complicated sensor network to be worn on human body. As such, wireless HAR, since emerging, has drawn significant attentions from both industry and academia. Current wireless HAR researches can be largely divided into two categories: device-based and device-free. Device-based approaches need wearing extra wireless devices (e.g., smartwatch, phone), which sometimes can be inconvenient, especially the devices usually need to be frequently recharged, making device-free HAR systems especially WiFi-based HAR solutions often a more pleasant choice. Recent advances of deep learning models can achieve end-to-end recognition, which further boost the HAR performance beyond hand-crafted features and traditional machine learning models, and gradually become the new mainstream due to the high recognition accuracy.
On the other hand, although it is relatively easy to achieve accurate recognition in a single domain (i.e., training and testing are in the same consistent WiFi environment), it is usually much harder to generalize a single-domain-trained model to other domains, especially as the WiFi-based HAR environments can change significantly.
Therefore, we develop a semi-supervised domain adaptation method for WiFi-based HAR (SSDA-HAR) called TOSS, and applies semi-supervised domain adaptation to WiFi-based HAR work for the first time. Besides, an uncertainty-based method is adopted to divide inner set so as to better benefit meta learner for fast adaptation on the source domain samples.
Methodology
We take CSI measurements as the original input data as well. CSI is a fine-grained measurement that provides abundant channel properties between the WiFi transmitter and receiver. As for device-free activity recognition tasks, CSI is sensitive enough to human actions. When human moves, the state of each channel changes.
To fully take advantage of the target domain samples, we propose to develop a target-oriented semi-supervised domain adaptation method. We denote the task whose samples are drawn from the same domain as homogeneous task, and the task whose samples are drawn from different domains as heterogeneous task. If there only exists one source domain, we just sample homogeneous tasks, otherwise we sample hybrid (i.e. homogeneous and heterogeneous) tasks.
Next we perform pseudo-label assignment on the labels. To dynamically balance soft and hard labels, we propose to leverage a dynamic entropy-based balance method to solve this problem.
To help TOSS learn useful knowledge from the target tasks more efficiently, we also optimize the meta learner so that it will not be biased to certain source task.
Achievement
In one-to-one single-shot domain adaptation, we can observe that our proposed TOSS is further enhanced beyond MAML and achieves the best overall performance. It is worth noting that TOSS also outperforms the state-of-the-art unsupervised domain adaptation model HDA and semi-supervised domain adaptation model PACL, which proves the superiority of the training principle of meta learning and our strategy in leveraging the target samples.
In multivariate single-shot In domain adaptation, the accuracy of TOSS also exceeds other existing models.
Extensive experiments demonstrated that our method can greatly outperform the state-of-the-art models, and can provide a promising solution for a series of real-scene problems.
