ICGNN Model for Fall Detection Using Inertial Sensors

Falls are a major cause of increased morbidity and mortality in the elderly. To solve the problem that it is challenging to detect human activity due to the fixed perspective of the camera, easy to be impeded and affected by the environment, a human fall detection method based on a wearable inertial sensor and convolutional recurrent neural network is proposed. First, multiple inertial sensors are used to capture the acceleration, angular velocity, and angle signals of the subject's fall and non-fall action. Then, the signal data is pre-processed and a data representation method suitable for multi-sensors is designed to realize the mapping between labels and signals. In this paper, an Inception-based convolution module and gate recurrent unit neural network (ICGNN) is proposed to extract multi-scale and temporal features of human actions from inertial signals and automatically classify fall actions. Experimental results show that the recognition rate of the model is high, and the accuracy can reach 93.36%, which has a good effect on the recognition of fall actions.