Interrelation Analysis for Interpersonal Behaviour Understanding in Social Context Kamrad Khoshhal Roudposhti * Jorge Dias * , ** * University of Coimbra, Portugal (e-mail: {kamrad,jorge}@isr.uc.pt). ** Khalifa University, UAE (e-mail: jorge.dias@kustar.ac.ae). Abstract: In this paper we study a probabilistic approach to characterize Interpersonal Behaviours (IBs) in a social concept by exploring the existent interrelation between body motion features. Human activities were explored in different level of complexities, such as social-based human activity. To bridge the existent big gap between human body motions and the IBs analysis, a set of proper dependencies definition between the features is vital. Inspired in the works of Alex Pentland and Rudolph Laban, we proposed a couple of layers of analysis. In the first layer, we analyse human body parts motions based on a known body motion descriptor, Laban Movement analysis (LMA). LMA composes a set of components which provides different types of human movement features. We investigated the interrelation between those LMA features of a couple of persons to provide a proper model to estimate the IBs in the second layer. To reach the goal, LMA components are used as body motion features. To computerize the model, Dynamic Bayesian Network (DBN) approach is used, because of its flexibility in development and implementation of the dependencies and interrelations. The results show the importance of the interrelations to have more accurate results of the IBs estimations. Keywords: Interrelation analysis, interpersonal behaviour analysis, social signals, Bayesian approach, Laban movement analysis. 1. INTRODUCTION People use their skill of body motions in communication to express better their points. In any human interaction, be- tween human body motions, there are several meaningful relations with respect to each others. Imagine two persons interact to each other, and each of them tries to respond other’s request. During those interactions we could see a relation between their body motions which assist us to re- alize the people and context situation even when we could not hear their conversation. Those features also are more reliable features to understand actual human behaviours, which Pentland call it “Honest signals ” (Pentland [2008]). In a social concept, it can be realized that each person is interested or influenced to communicate with others by observing their body motions. For instance in a TV show program, the showman use body motions too much to attract audiences, but a newscaster is in opposite situation. However it also depends on the person’s attitude, culture, etc. In this paper, we intend to explore in the existent rela- tions between people body parts motions, which plays an important role, to analyse the Interpersonal Behaviours (IBs). To implement the idea we propose Laban Movement analysis (LMA), which is a known body motion descriptor as a mid-level features. LMA has several components, which were investigated, analysed and modeled (Zhao and Badler [2005], Rett [2008], Khoshhal and et al. [2011b]), to describe human movements with several symbols. Those descriptions are useful not only on the modeling of complex human activities, but also for finding out and analysing the existent relationships between body parts motions in different IBs. Dynamic Bayesian Network (DBN) is the proper approach to have the flexibility to perform those dependencies (relations). In the last decade, researchers were interested to under- stand human behaviours in different applications, such as surveillance, security and social systems, thus many approaches were introduced. Maja Pantic’s group cate- gorized those approaches based on the existent types of observation data; facial expression, voice, and body motion (Pantic et al. [2006]). Each of those approaches has own advantages and disadvantages. In this paper, we attempt to rely just on body motion types features to analyse human behaviours in social aspect. In many applications, having an acceptable visibility of face image and voice data is complicated. For instance in many public places which usually there are several cameras around, collecting body motion type features are more proper than face and voice ones. Most of the attempts in this kind of applications just used motion-based features from the human as a blob, which cannot be useful for analysing of complex types of human activities such as handshaking or even more complex such as mimicry. Body parts motions are very informative for analysing human activities which are not applicable even by other types of features. By progressing of existent techniques about 3D reconstruction of human body, such as (Aliakbarpour and