Cognitive perceptual models in machine learning of human activities

Project start: 

15 October 2012

iCareNet Fellow: 


Abstract of the Projects

Most approaches aiming to models the activity of humans are typically inefficient due to their inability to meet the inherent variability of the human behaviour. Furthermore, a real deployment of Healthcare, Wellness and Assisted Living applications is limited by the inefficient need of the activity recognition system to acquire sensors data by a continuous streaming. This project will investigate advances in cognitive theories to develop activity recognition and behaviour inference models that can cope with changing concepts and allow dynamic reconfiguration  of sensor networks at run time.

Research Directions


Based on the concept of Activity-Event Detector shown in Fig.1, distributed artificial intelligence techniques will be investigated that will enable the automatic discovery of atomic structures constituting complex activities. The techniques should be adapted for enabling the automatic reconfiguration of the AED digraph structures, as shown in Fig. 2, ensuring efficiency in the power management of the sensors.

Machine Learning Techniques based of Hierarchical Multiple Classifiers Systems will be considered in the research. Alternative solutions based on generative models will be also taken into account, e.g. Restricted Boltzmann Machines. At the same time, approximated data mining and knowledge discovering algorithms will be considered for continuous learning of changing concepts in human activities in order to understand the behaviour of humans in smart environments. 



1] Amft O., Lombriser C., Modelling of distributed activity recognition in the home environment, IEEE EMBS, 2011

[2] Lombriser, C., O. Amft, P. Zappi, L. Benini, and G. Tröster, "Benefits of Dynamically Reconfigurable Activity Recognition in Distributed Sensing Environments", Activity Recognition in Pervasive Intelligent Environment, vol. 4: World Scientific Publishing Co.. 2010