Identification of Gait Events Combining Bayesian Hidden Markov Models and Linear Regression

The Hidden Markov Model is a probabilistic time- series model that has recently found application in human motion analysis. HMMs are usually fit directly to time-series data obtained from motion capture systems, using Gaussian ob- servation models and the Expectation Maximization algorithm. The boundaries of the segmentation induced by the HMM are somewhat arbitrary, because the motion capture data usually consists of smooth trajectories. When a-priori segmentation is available, like in the case of clinically defined events in human gait, biasing the HMM parameters towards this prior knowledge is crucial to obtain a segmentation that is clinically relevant. To achieve this goal, we propose the combination of a fully Bayesian HMM with a sliding-window polynomial fit pre-processing step. In the context of automatic segmentation of gait time-series, we show how the proposed approach allows to better exploit a-priori segmentation, and to learn a set of motion primitives that improve the segmentation performances over classical HMMs.status: publishe