Abstract
With the pervasive use of smartphones, which contain numerous sensors, data for modeling human activity is readily available. Human activity recognition is an important area of research because it can be used in context-aware applications. It has significant influence in many other research areas and applications including healthcare, assisted living, personal fitness, and entertainment. There has been a widespread use of machine learning techniques in wearable and smartphone based human activity recognition. Despite being an active area of research for more than a decade, most of the existing approaches require extensive computation to extract feature, train model, and recognize activities. This study presents a computationally efficient smartphone based human activity recognizer, based on dynamical systems and chaos theory. A reconstructed phase space is formed from the accelerometer sensor data using time-delay embedding. A single accelerometer axis is used to reduce memory and computational complexity. A Gaussian mixture model is learned on the reconstructed phase space. A maximum likelihood classifier uses the Gaussian mixture model to classify ten different human activities and a baseline. One public and one collected dataset were used to validate the proposed approach. Data was collected from ten subjects. The public dataset contains data from 30 subjects. Out-of-sample experimental results show that the proposed approach is able to recognize human activities from smartphones' one-axis raw accelerometer sensor data. The proposed approach achieved 100% accuracy for individual models across all activities and datasets. The proposed research requires 3 to 7 times less amount of data than the existing approaches to classify activities. It also requires 3 to 4 times less amount of time to build reconstructed phase space compare to time and frequency domain features. A comparative evaluation is also presented to compare proposed approach with the state-of-the-art works.