We aimed to acquire robust feature representations for action recognition across different environments using adversarial training methods, as exemplified by Generative Adversarial Networks (GAN: Goodfellow2014). In the following paper, we addressed the problem of potentially varying sampling frequencies across different measurement devices by proposing the following two methods, achieving improved estimation accuracy.

Among action recognition methods, discussions regarding deep learning have been active in recent years. However, unlike the image field, it is still in the early stages of research. For instance, various CNN-based representation learning methods developed for images have been repurposed for action recognition, but these methods strongly emphasize the local neighborhood information characteristic of images. It would be desirable to develop layers that can acquire appropriate feature representations for low-sampling-rate waveform data such as sensor data.

Under the condition that observed accelerometer waveform data differs depending on the action, it is common to automatically classify this using machine learning and other methods. There are various research tasks, but the following two are considered representative.