This PhD, proposed by the EXPRESSION team at IRISA, is focused on the general framework of "anomaly" detection in multicanal sequences. An "anomaly" is charaterized by the existence of foreign elements to a normal situation in a precise context. These sequences can concern both temporally and spatially changing data: for instance, vocal and video recordings, but also system call sequences on a host machine of a network. The goal is to generate models of abnormal behaviors through machine learning methods.
This study can occur naturally in the particular context of the detection of an abnormal behavior of a human being from the facial movements and the vocal signal. We think of extreme stress situations for airplane pilots or machine operators, for example. One can also think of the detection of hostile behaviors by observing a vocal statement and a neutral face in a situation where the speech should be relaxed and expressive. This study could also interest applications in the medical field, as for example, the detection of abnormal behaviors due to psychic handicaps such as autism. Finally, the evaluation of the falsification of a detection system is possible.
We aimed at developing a system capable of detecting abnormal behaviors by the analysis of records of concrete situations. The thesis will then explore several issues as collect, segment and annotate multimodal data; Identification of descriptors enabling the description of abnormality; Development of dedicated machine learning approaches for abnormality detection; Development of a decision system.
Please send a CV, application and reference letters, academic results of previous diploma to all the contacts: Arnaud Delhay, Pierre-François Marteau and Damien Lolive BEFORE April 10th 2018.