An efficient 3D convolutional neural network with informative 3D volumes for human activity recognition using wearable sensors

Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT) are two popular transforms which can be used to find time-frequency representations. By using them, one-dimensional signals acquired from different axes or sensors are mapped to time-frequency representations. These representations can construct 3D volumes which include time-frequency information of signals. Recently, the advantage of 3D convolutional neural networks (3D-CNNs) for video classification causes to incorporate them with the 3D volumes. Based on this opinion, a novel method composed of two basic methods is proposed in this paper. The magnitude of the STFT and the CWT are used for constructing 3D volumes in basic methods. Also, a developed 3D-CNN is applied for classifying. Two streams of these 3D volumes are fused in the proposed method. It attains the accuracies of 96.61%, 97.77%, 99.65% and 98.32% for UCI HAR, MOTIONSENSE, MHEALTH and WISDM datasets, respectively. Achieved results demonstrate the superiority of the proposed method compared with state-of-the-art approaches.