Adaptation and Selection techniques based on Deep Learning for Human Activity Recognition using Inertial Sensors

Deep learning techniques have been widely applied to Human Activity Recognition (HAR), but a specific challenge appears when HAR systems are trained and tested with different subjects. Each user presents different patterns when performing several physical activities, so HAR systems should adapt the activity models trained with some users’ data to new subjects. This paper describes and evaluates several techniques based on deep learning algorithms for adapting and selecting the training data used to generate a HAR system using accelerometer recordings. This paper proposes two alternatives to adapt and select the training data: autoencoders and Generative Adversarial Networks (GANs). Both alternatives are based on deep neural networks including convolutional layers for feature extraction and fully-connected layers for classification. Fast Fourier Transform (FFT) is used as a transformation of acceleration data to provide an appropriate input data to the deep neural network. This study has used acceleration recordings from hand, chest and ankle sensors included in the PAMAP2 dataset. This is a public dataset including recordings from nine subjects while performing 12 activities such as walking, running, sitting, ascending stairs or ironing. The evaluation has been performed using a Leave-One-Subject-Out cross-validation: all recordings from a subject are used as testing subset and recordings from the rest subjects are used as training subset. The obtained results suggest that strategies using autoencoders to adapt training data to test data improve the general performance. Moreover, training data selection algorithms with autoencoders also provide improvements. The GAN approach, using the discriminator module, provides a significant improvement in adaptation experiments.