Abstract
Activities recognition using a wearable device is a very popular research field. Among all wearable sensors, the accelerometer is one of the most common sensors due to its versatility and relative ease of use. This paper proposes a novel method for activity recognition based on a single accelerometer. To process the activity information from accelerometer data, two kinds of signal features are extracted. Firstly, five features including the mean, the standard deviation, the entropy, the energy and the correlation are calculated. Then a method called empirical mode decomposition (EMD) is used for the feature extraction since accelerometer data are non-linear and non-stationary. Several time series named intrinsic mode functions (IMFs) can be obtained after the EMD. Additional features will be added by computing the mean and standard deviation of first three IMFs. A classifier called Adaboost is adopted for the final activities recognition. In the experiments, a single sensor is separately positioned in the waist, left thigh, right ankle and right arm. Results show that the classification accuracy is 94.69%, 86.53%, 91.84% and 92.65%, respectively. These relatively high performances demonstrate that activities can be detected irrespective of the position by reducing problems such as the movement constrain and discomfort.