In recent years, research on human behavior recognition using Wi-Fi has been active, with the following advantages: Wi-Fi has less privacy concerns than video, does not require maintenance such as recharging, and can be used with existing Wi-Fi facilities, thus reducing installation costs. Wi-Fi, however, is not suitable for all environments. However, Wi-Fi is susceptible to environmental influences, and in many cases, learning is required for each environment. In this study, we use an extended Autoencoder to extract environment-independent features, which can then be used for environment-independent activity recognition. The extended Autoencoder learns to restore the same action data acquired in a different environment when data from one environment is input. In addition, multiple decoders are prepared for each environment for each Encoder, and each decoder is trained to restore data from a different environment. This is expected to eliminate environmental factors in the Encoder part and extract only features related to behavior. Using the extracted features, deep learning such as FNN (Feedforward Neural Network) is used to perform environment-independent activity recognition.
