MEMS-based Human Activity Recognition Using Smartphone

Data mining is one hot orientation in today's research field. Human activity recognition is meaningful in our daily living and is a significant aspect in data mining. Most previously research is almost based on tri-axial accelerometer. This paper presents a novel method to collect data from both accelerometer and gyroscope using smartphone. Our daily activities including Walking, Running, Upstairs, Downstairs, Standing, Sitting and Cycling, a total of seven categories are classified. The raw data from MEMS are recorded by smartphone according to different daily activities. To improve the accuracy of classification for daily activities, this paper combines time-series features with wavelet coefficients to extract features. To recognize these activities, the support vector machine is used to finish this work. Besides, we compare the accuracy with other machine learning methods, such as k-nearest neighbor algorithm and neural network or decision tree. The result indicates that our method can achieve nearly 96% classification accuracy for the seven kinds of daily activities.