|Probabilistic-based Human Behaviour Analysis using Hierarchical Framework|
Seminar by Kamrad Roudposhti (preparation for Ph.D. defense)
(Supervisor: Prof. Jorge Dias)
Synopsis: Recently by progressing technology and infrastructures, monitoring and understanding human behaviour and activity is going to be more interesting in various applications. Exploring through human body parts motions to analyse human behaviours in different contexts, is the aim of this PhD research. The features of human movements are less restricted than the other possible features (e.g. facial expressions and voice) in many real world applications, however they are more complicated to be analysed. Due to the large dimensions of body parts movements and the complex dynamics and dependencies between them, high computational processing resources are needed to analyse and estimate human behaviours. Firstly, a hierarchical framework is proposed to reduce the complexity of the process in different layers. Secondly, a well-known human movement descriptor, Laban Movement Analysis (LMA), which provides different types of needed features in ?ve components, is proposed. The LMA components prepare minimum needed features that can assist us to analyse any kind of human activities, and ?ll the gap between the Low Level Features (LLFs) and human movements analysis. Finally, for modeling the framework, Bayesian-based approaches (Bayesian Network (BN), Dynamic BN (DBN), and Hidden Markov Model (HMM)), are defined to deal with the uncertain data, to apply learning processes using small data, to fuse different types of features (in frequency and spatial domains) and to have enough ?exibility for modeling the different dependencies between different features and layers. In this study we explored several different human activities and behaviours though the framework, namely; body parts movements, human individual analysis, human-object interaction, human-human interaction, interpersonal behaviour, and social role of people. The mentioned framework is modeled and constructed in a bottom-up strategy. During the process, different approaches are proposed to solve the problems. From the lowest level of analysis, the study is divided into two different domains; frequency and spatial. This study is performed to estimate LMA components, which provide enough knowledge in body motion level, to be able to explore more high level of human activities analysis. For instance; Effort component, which explains human body movement dynamics, is modeled in frequency domain, and Shape component, which explains human body shape deformation in 3D space during any movement, is modeled by spatial based features. Based on the mentioned components, individual human action level is modeled. In the next step to analyse human-object and human-human interaction analysis (context-based), each individual human actions and the relations between them, are needed. Those relations are modeled by inspiration of Relationship Component which is one of the less explored LMA component. Thus we were able to model human-object and human-human interactions by modeling Relationship components through the framework to explore human activities with respect to the context (i.e., scene understanding). With respect to the proposed framework, human activities in social context are explored by information of the LMA level. The system was enough ?exible to provide the complex existent dependencies between di?erent features to estimate body-motion based interpersonal behaviors and a social role, inspired by Alex Pentland investigation in ”Honest signals” book. The hierarchical framework presents many capabilities such as; ?exibility of modeling, generalizing to di?erent related applications, extendability by progressing the sensory technology, dealing with uncertainly in all level of analysis, and providing semantic-based information for all layers of analysis. The proposed framework provides an automatic monitoring human behaviour system which is very highly interested application in almost everywhere that people are involved, such as; clinical study, security system, elder-care, surveillance system, sport training, virtual reality, choreography, etc. In the experimental process, a motion tracker suit which provides 3D position of human body parts in maximum 120 Hz resolution, is used. To prepare a dataset, several people dressed the suit and performed the defined activities. The attached sensors of the suit for each record trial, need to be calibrated. It means the data for each person in different trials can be different with respect to the calibration process. The obtained results in the each step, present the capability of the mentioned framework in different level of human movement activity analysis. Furthermore, a descriptive global framework to explore and estimate various level of human body-motion based activities, is proposed.