Optimising recognition rates for subject independent gait pattern classification

This paper describes a study which was carried out to determine an optimally performing classification algorithm for the problem of subject independent gait pattern classification. The study utilised a frequency domain based feature vector based on the concept of cepstral coefficients whose generation methodology was optimised in terms of overall system recognition rates. The performance of a number of both linear and nonlinear classification algorithms including neural network and Support Vector Machines was examined. An optimal recognition rate of 78.4±3.2% was achieved using a "one-versus-all" MLP classier applied to a previously unseen test database of 12 subjects completing ten repetitions of five different human gait patterns including walking on level surfaces, walking up and down stairs and walking up and down ramps.