Feature Selection for Activity Recognition from Smartphone Accelerometer Data

We use the public Human Activity Recognition Using Smartphones (HARUS) data-set to investigate and identify the most informative features for determining the physical activity performed by a user based on smartphone accelerometer and gyroscope data. The HARUS data-set includes 561 time domain and frequency domain features extracted from sensor readings collected from a smartphone carried by 30 users while performing specific activities. We compare the performance of a decision tree, support vector machines, Naive Bayes, multilayer perceptron, and bagging. We report the various classification performances of these algorithms for subject independent cases. Our results show that bagging and the multilayer perceptron achieve the highest classification accuracies across all feature sets. In addition, the signal from gravity contains the most information for classification of activities in the HARUS data-set.