Abstract
Biological motion has a lot of information, both physical and cognitive. Studies have shown that it is possible to determine a person’s gender, emotion and even identity. These characteristics are accessible from information on the dynamics of the movement of polyarticulated bodies (e.g. the movement of the articulation points). Understanding and interpreting a person’s behavior and state are abilities related to empathy. It is a faculty common to all mammals and is based on certain neural systems including, among others, mirror neurons. Given that empathy is an important part of social interactions in humans and more generally in animals, we can ask ourselves how our relationship with robots can be inspired by it.
This leads us to the following problem: can robotic assistance to people use the interpretation of human movement, rich in physical and cognitive information, as a modality to improve the Human-Robot Interaction?
To answer this question, we are working on observation tools and on a method of motion analysis that can be used in real time by a robotic system.

Initially, we worked on the observation tools of human movement. Our objectives of robotic assistance in an ecological environment require the installation of sensors that affect the person’s actions as little as possible. We have therefore chosen to study the Microsoft Kinect sensor which is an accessible depth sensor allowing to recover the Cartesian positions of the joints and extremities of the body. However, this type of sensor is subject to measurement noise that would prevent a fine analysis of the movement. We have therefore developed two methods to improve the measurement of this sensor based on the Extended Kalman Filter (EKF): an anthropometrically constrained EKF and a sensor fusion EKF. We have done the first study on the 2nd generation Kinect and the second on the 2nd and 3rd generations, allowing to highlight the differences between these two sensors.

In a second time, we were interested in motion analysis methods and more specifically in the problem of Inverse Optimal Control (IOC). The objective of IOC is to identify the weights associated with a set of cost functions to be optimized to generate a given trajectory. In this thesis, we seek to analyze in real time human motion trajectories whose measurements, coming from sensors, are noisy. We have studied the reliability of the IOC resolution method called Approached, as a function of the measurement noise. We also provide an original approach to the IOC that poses a new view of the optimality of trajectories and allows us to introduce the concepts of Singularity Curves and Projection. We show in this paper tools to better understand and take into account the robustness issues of IOC.