Latent force models for human action recognition

Abstract

Human action recognition is a key process for robots when targeting natural and effective interactions with humans. Such systems need solving the challenging task of designing robust algorithms handling intra and inter-personal variability: for a given action, people do never reproduce the same movements, preventing from having stable and reliable models for recognition. In our work, we use the latent force model (LFM [2]) to introduce mechanistic criteria in explaining the time series describing human actions in terms actual forces. According to LFM’s, the human body can be seen as a dynamic system driven by latent forces. In addition, the hidden structure of these forces can be captured through Gaussian processes (GP) modeling. Accordingly, regression processes are able to give suitable models for both classification and prediction. We applied this formalism to daily life actions recognition and tested it successfully on a collection of real activities. The obtained results show the effectiveness of the approach. We discuss also our future developments in addressing intention recognition, which can be seen as the early detection facet of human activities recognition.