Development of human activity recognition component

The task of recognizing and predicting human daily activities is a trending topic nowadays, and a lot of research has been developed around it, accompanied with the creation of algorithms that achieve state-of-the-art results on different human activity data sets: CAD-60 or CAD-120, UTKinect-Action, Florence3D-Action data sets, etc. The application of a machine that can detect a person’s actions is broad. It has been developed for gaming, Human-Computer interac-tion or Active and Assistive Living.

For the development of RoboComp's framework, my work will be based on the article of Premebida, Souza and Faria (2017) where the proposed algorithm reached an accuracy of 94.74% and a recall of 94.74 % on CAD-60, which is considered as a state-of-the-art result. These articles are mainly based on Dynamic Bayesian Networks and a Dynamic Bayesian Mixture Model (Faria,Premebida,Nunes(2014)). Other works will be taken in account to take advantage of what we have available, combining machine learning approaches with mathematics to get robust results, such as the introduction of Partial Differential Equations or Lie groups, explained, for example, in Vemulapalli et al. (2014)