|
Doctoral researcher Mechanical Engineering Department Katholieke Universiteit Leuven, Belgium |
ContactMailing addressMechanical Engineering Department - PMA Division Katholieke Universiteit Leuven Celestijnenlaan 300B postbusnr.: 2420 B-3001 Leuven, Belgium Contact info enrico_dot_dilello_at_mech_dot_kuleuven_dot_be Phone: ++32 16 32 80 58 Office: 01.053 |
Research
Probabilistic Graphical Models for multi-dimensional time-series recognition, applied to human motion estimation and classification and robotic assembly process monitoring and learning. The main challenge of this research is to apply Probabilistic Graphical Models ( Dynamic Bayesian Network as a first approach) to classification of time series. In particular, my research will focus on the recognition of two categories of input: - Human motion data In collaboration with the Department of Revalidation Sciences (FaBeR), I plan to apply both static and dynamic Bayesian Networks for the classification of 3D gait analysis data acquired via a Vicon system. The goal is to support doctor in the identification of clynically relevant motion patterns in the gait of children affected by cerebral palsy. The proposed approach is to develop a DBN that can be trained with data labelled from experts, and then classify the data acquired by the motion capture system to one of the expert defined categories, providing also confidence values. I am currently investigating the use of Hidden Markov Models, from the standard version to the more complex Hierarchycal and Layered-HMM ones, for the motion analysis. A more general DBN, although, will be needed to embed the a-priori knowledge provided by expert and obtain classification performances comparable with the ones obtained by doctors and by current state-of-the-art approaches. Another application i am currently interested in is human motion capturing with the Kinect sensor. In particular, I am exploring how the on-line classification of human motions can be used to improve the markerless estimation of the current human pose. Recently, prof. De Schutter proposed a representation of rigid body 6 DOF motion that is invariant with respect to the observation point. I am working in combining this representation with standard DBN approaches for motion classification for robotics application. - Force sensor output The output of 6D force sensor can be interpreted to perform "smart monitoring" of an assembly task. In industrial robotics application, a manipulation or assembly task is usually encoded in a FSM, which transitions are usually triggered by time-based "event". By learning probabilistic models of this "events" from temporal patterns in the force sensor output we aim at improving robustness, failure recovery ability and speed. The research is promoted by Prof. H. Bruyninckx at the Katholieke Universiteit Leuven, Belgium, in collaboration with both the Department of Mechanical Engineering (PMA) and the Department of Revalidation Sciences (FaBeR). |
PublicationsJournalsConferences
Workshops
|
Software
|
LinksGroup: People:
|
Projects
|
|
|
|
|
|
|