Long-Term Activity Recognition from Accelerometer Data

In the last years, simple activity recognition through wearable sensors has been achieved successfully, however complex activity recognition is still challenging. Simple activities may last just a few seconds, e.g., walking, running, resting, etc. whereas complex activities involve a combination of the former and they may last from a few minutes to several hours. In this work long-term activity recognition is performed and modeled as a distribution of simple activities represented as a histogram. For the experiments, the raw histograms were used for the recognition task and then we added an additional step which consists of extracting features over the histogram and applying a simple threshold to reduce noise. This additional step resulted in an increase on the classification accuracy.