Hidden Markov models applied onto gait classification

This paper is about recognition of different gait conditions from body-worn sensor data. Our sensor, located at subject's shank, is a combination of a 3-D accelerometer and a 3-D magnetometer. Stride detection method relies on the use of the sole magnetometer readings. Feature extraction combines both modalities in an original manner and spatial, temporal, and angular parameters are extracted for subsequent classification. Hidden Markov models are employed to identify the types of gait being performed. Different feature modelizations are typically considered with the use of Gaussian mixture laws. This paper analyses which stride feature sets are the most significant and what could be the minimal number of training sequences for best classification scores. Classification performances above 90% are demonstrated.