In the earlier studies we have developed activity recognition algorithms which are based on features calculated from data of 3D accelerometer sensor placed on the hip, close to the centre of mass. In the development subjects have been young adults. Now we study if the input features of the algorithm are generalized for different set-ups; for older adults and when the sensor is worn as a necklace. From the 3D accelerometer resultant magnitude the following features were calculated for each second: spectral entropy, peak frequency, power and range. The frequency domain features behaved in a relatively stable manner in the set-ups but the time domain features differed significantly from statistical and algorithm perspective between the set-ups. By developing time domain features to be more inter-individual independent would be beneficial for activity recognition algorithms.
展开▼
