Dynamic Activity Recognition using Smartphone Sensor Data

Smartphones equipped with various sensors provide sufficient sensor data and computation power to enable daily activity detection for applications such as u-healthcare, elderly monitoring, sports coaching and entertainment. Instead of applying multiple sensor devices, as observed in many previous investigations, this work proposes the use of a smartphone with its built-in accelerometer as an unobtrusive sensor device for real time activity recognition of basic daily activities. The proposal is tested experimentally through evaluations on real data collected from 50 participants. A prototype application is developed to demonstrate and evaluate the selected classification methods for the designated recognition tasks. The results indicates that the J48 classifier using a window size of 512 samples with 50% overlapping obtained the highest accuracy (i.e., up to 96.02%). To measure the actual classification accuracy, a 5x10-fold cross validation with different random seeds was performed on the dataset using WEKA. Finally, to determine whether a classifier is superior to another, 5x2 fold cross validation along with a paired t-test was subsequently performed on the results using J48 as the baseline scheme with the other classification algorithms being compared to it. A value of p<0.05 was considered statistically significant.