Artificial neural networks in accelerometer-based human activity recognition

This paper presents a study aimed to assess applicability of artificial neural networks (ANNs) in human activity recognition from simple features derived from accelerometric signals. Secondary goal was to select the most descriptive signal features and sensor locations to be used as inputs to ANNs. Five triaxial accelerometers were attached to human body in the following places: one at back, two at waist laterally and two at both ankles. The set of activities to be recognized was established to include the most often performed actions in home environment. In total 25 subjects performed a set of predefined actions like walking, going up and down the stairs, sitting down and standing up from a chair. Acquired signals were divided into 0.5s time windows by a label defining the action performed. Several statistical signal features were calculated and used to train ANNs. Learning and testing were performed on separate data sets. Analysis using Fisher Linear Discriminant showed that despite the fact that some of the calculated values play a significant role in the distinction between similar activities, none of the features or sensors could be omitted in the recognition of the activities considered in the study. Accuracy of 97% has been achieved for discriminating sitting and walking, 89% for standing, 72-75% for walking the stairs. Transient actions like standing up and sitting down have been detected with accuracy 56% and 38%, respectively. Even though there are studies declaring higher accuracy, none of them considered a set of activities analyzed in this research.