Using Accelerometers for Physical Actions Recognition by a Neural Fuzzy Network

Abstract

Triaxial accelerometers were employed to monitor human actions under various conditions. This study aimed to determine an optimum classification scheme and sensor placement positions for recognizing different types of physical action. Three triaxial accelerometers were placed on the chest, waist, and thigh, and their abilities to recognize the three actions of walking, sitting down, and falling were determined. The features of the resultant triaxial signals from each accelerometer were extracted by an autoregression (AR) model. A self-constructing neural fuzzy inference network (SONFIN) was used to recognize the three actions. The performance of the SONFIN was assessed based on statistical parameters, sensitivity, specificity, and total classification accuracy. The results show that the SONFIN provided a stability total classification accuracy of 96.3% and 88.7% for the training and testing data, when the parameters of the 60-order AR model were used as the input feature vector, and the accelerometer was placed anywhere on the abdomen. Seven elderly individuals performing the three basic actions had 80.4% confirmation for the testing data.