Detecting and exploiting periodicity in activity classification

The technology for activity classification presents new opportunities for control and monitoring of serious games players. Other than for step detection, human activity classification is normally undertaken by calculating features from a fixed interval length of sensor data and comparing them to values expected from a range of activities. It was observed that many human activities, especially vigorous activities, are cyclic. This paper tested the hypothesis that, if features are calculated over an integer number of cycles of a cyclic activity, more accurate activity classifications will be achieved. An algorithm was developed that determines whether an activity is cyclic and if so, identifies the cycles and calculates the feature set over an integer multiple of cycles. If the activity is determined to be non-cyclic, the features are calculated over a fixed time window. The hypothesis that more accurate activity classifications will be achieved was confirmed with a pairwise t-test at the 99% significance level. Knowledge that a cyclic activity is taking place is already informative and can enable generation of features from individual cycles. One of these, cycle length, was added to the feature set which improved recognition rates for activities where the cycle length varied greatly from other activities. For example, brushing teeth improved from 39% with a fixed window to 54% for an adaptive window and 74% for an adaptive window with cycle length. An overall improvement in activity classification success rate was found for this adaptive windowing method compared to a fixed window approach with overall success rates of 54% for a fixed window, 62% for an adaptive window and 64% for an adaptive window with the additional feature.