Langevin Diffusion on Infinite Dimensional Exponential Family for Human Gait Simulation

Music City Center 

We consider a high dimensional and high resolution time series dataset of human gait for the purpose of predicting and preventing falls. This data was collected via motion capture of 13 individuals walking across a variety of artificial terrains. The angle of 9 joints were continuously measured through a total of 15000 strides. Our goal is to develop an approach for modeling the individual distribution of gait patterns, so that deviations from that individual distribution can serve as warning signs of increased fall risk.

We propose a Langevin Diffusion Model with a nonparametric stationary distribution, which we model using an Infinite Dimensional Exponential Family. This model can flexibly estimate individual-level stationary distributions of gait style and also permits detection of departures from an individual's movement profile. We use this model to simulate synthetic gait data, and also to develop methods for identifying In turn, this prediction apparatus will enable simulation of gait data for engineering applications and potentially improve health outcomes through fall prevention.