Accelerometer-based activity recognition on a mobile phone using cepstral features and quantized gmms

The use of cepstral coefficients derived from a filter bank with logarithmically spaced band center frequencies and Gaussian mixture models (GMMs) with quantized parameters (qGMMs) are proposed for accelerometer-based activity recognition of mobile phone users. The use of a filter bank with logarithmically spaced band center frequencies is shown to yield better results than the use of a filter bank with linear spacing between band center frequencies. GMMs and qGMMs are shown to achieve similar recognition accuracies. However, the computation time using qGMMs is shown to be either at the same level or faster when compared to GMMs, depending on model complexity. Using the proposed approach, we achieve an accuracy of 72.6% and 91.3% on two recognition tasks with seven and five activities, respectively.