Section: New Results

Filtering method for human motion analysis

We have developed a series of filters to estimate the states of a dynamic system from a series of incomplete or noisy measurements for the analysis of human motion. They are also used for data fusion or for filtering noisy data from a model, especially for a Kinect and Orbbec sensor. In our case, we first developed an extended Kalman filter [13] that we improved to take into account the singularities of representations of the human kinematic module, the estimation of users' physiological parameters as well as the calibration of measurement systems. In addition, different strategies have been implemented to ensure the real-time operation of the filter, and the addition of joint constraints to improve the accuracy of the results.

In a second step, we implemented an interesting alternative technique for filtering time series. It consists of performing singular spectrum analysis. Due to the multidimensional nature of the type of data we use a specific version of this technique called Multivariate or Multidimensional Singular Spectrum Analysis (MSSA) [19].

This technique is based on a method called decomposition into main components which aims to compress the data both on their temporal and physical dimensions. Excellent results have been obtained.