Human Activity Recognition Based on Extraction of Multi-feature Using Inertial Sensors

Recently, deep neural network has been widely used in the field of human activity recognition (HAR) and shows good performance. Feature extraction of activity signals is one of the keys of HAR. However, the features extracted by some existing methods are either incomplete or redundant. To obtain better effect of HAR, the extraction of multiple features is studied. In this paper, we proposed a hybrid deep neural network (DNN) based on a three-layer CNN network, a BiLSTM layer, and channel attention mechanism to extract the deep features. A channel attention method is proposed to adjust the stacked CNN network. A handcraft features extraction module is added to the proposed DNN to extract manual features of sensor data in time domain and frequency domain. The deep features and handcraft features are fused by the stacked dense layer and the HAR is completed by softmax layer. Finally, a small self-made IMU circuit was placed below the experimenter’s knee to collect data of human lower limb activity. The proposed model is validated through our dataset and public dataset: WISDM. The experimental results show that our proposed method has a higher accuracy compared with most related methods.