Human Activity Recognition Using Multinomial Logistic Regression

Abstract

Activity monitoring is one of the predominant concerns of the elderly living at home. For example, serious complications or even death can be avoided if the elder is helped shortly after a serious fall. Several methods have been proposed to recognize human activity. Most of them can be categorized in three general groups: (1) vision-based methods; (2) wearable sensors; and (3) vibration-based methods. Despite the advantages of vision-based methods and wearable sensors, privacy concerns in the first type, and compliance challenges in the second group motivate the health care industry to pay attention to other types of environmental monitoring techniques such as vibration-based methods. Vibration-based methods use accelerometers placed at the floor of the patient’s dwelling. These methods are appealing because accelerometers are not expensive and installation is easy. The classification of events is critical in vibration-based methods. This paper proposes a classification algorithm using Multinomial Logistic Regression (MLR). A data set of different human activities is experimentally obtained. Some signals are used to train the classification algorithm while the remaining records are used to test the proposed technique. The signal “peak” and a “proposed time index” are used for classification. Principal component analysis is used to reduce the number of parameters considered for MLR given that each event was captured by four sensors. The main advantage of this classification algorithm is its probabilistic nature. Results show that in most cases, the event was correctly identified by the group having the highest probability. However, other signal characteristics should be explored to improve the performance of the proposed technique.