A study in human activity recognition
Niazi, Anzah Hayat Khan
MetadataShow full item record
A three-stage hierarchical classifier, using Random Forests, was constructed to classify 23 different physical activities of various types. This classifier was built using triaxial accelerometer data from 77 subjects collected during trials in Phoenix, Arizona. The activities were hierarchically divided and five Random Forest classifiers were trained for each level. The classifier performed well compared to similar classification studies in this domain, achieving 94% for activity groups and 87% at the individual activity level. Furthermore, the effect of sampling rate and window size on activity recognition was also analyzed. Window size and sampling rate were varied, and a two-way weighted least squares analysis of variance was carried out. This analysis was carried out across a variety of activity types and demographic features. It was found that data collected at 50Hz, using 10 second windows performed statistically better than other data. There is, however, some statistical margin to allow for lower sampling rates and window sizes to be used without a significant reduction in classifier performance.