Human activity recognition by body-worn sensor data using bi-directional generative adversarial networks and frequency analysis techniques

Existing datasets used for human activity recognition (HAR) usually suffer from limitations in terms of size variation and distribution of activity classes, which can impair the generalizability and robustness of the trained model, especially in the case of activity classes with minority data. This paper proposes an architecture utilizing bi-directional generative adversarial networks (Bi-GANs) beside fast Fourier transform, which stacks the 1-D accelerometer signals as m frequency bins × frames × 3 orientations and produces an RGB-based features pattern. The extracted patterns allow the 2D-CNN-based Bi-GAN architecture to learn the accelerometer signals' cross-axis relationships, which served as input for training a deep learning model for activity recognition equipped with a fuzzy inference dense layer. Also, we have used Hidden Markov Models (HMMs) for post-processing the classifier's output, which integrates the window-level decision in more extended periods, obtaining a significant performance improvement. The proposed method examined and conducted on MobiAct, Up-Fall, Opportunity, and WISDM datasets with different rates of augmentation, reaching to the 99.7%, 99.96%, 86.8%, and 99.12% rate of accuracy, respectively.