Seminar by Luis Santos (preparation for Ph.D. defense)
(Supervisor: Prof. Jorge Dias)
This work presents a set of novel methodologies and concepts that enable robots to comprehensively interpret human motion and extend such knowledge via hierarchical analysis of different types of information. The proposed methods in this thesis address the following main three topics: (1) Defining a model which can robustly infer different types of information from human motion, using a general grounding language; (2) Encoding the unique expressive properties of each person’s motion, so as to develop action invariant motion signatures towards a person recognition framework; (3) Develop a system’s action memory, which can store and retrieve action generalized information towards incrementally learn new actions and executing them to solve a task in has respectively. We start by presenting an innovative approach to hierarchical analysis of human motion, based on a descriptive motion language, Laban Movement Analysis. This allows a system to infer multiple levels of information, from dynamic characteristics to intentions or behaviour patterns, by observing the 3D trajectories, generated from a given motion instance. Then, we exploit the outcome of Laban qualities classification into encoding this information to develop individual motion profiles. Such characteristics are then applied to develop a Bayesian-based action invariant person recognition framework. The two aforementioned techniques are then integrated and adapted to develop an intelligent video-surveillance framework, showing to be capable of robustly recognize actions and person identities. The last part of this work focuses on developing a set of cognitive skills, allowing the system to build its own memory, by either learning new actions or incrementally fuse newly performed actions to existing knowledge. All methods have been developed using probabilistic learning and inference, more specifically, Bayesian methodologies. They have been implemented and thoroughly evaluated using cross-validation procedures and different kinds of experimental scenarios so as to allow withdrawing conclusions based on produced evidences. Results demonstrate a highly robust and precise framework, whose main characteristics are flexibility, scalability and adaptability, showing to be useful to increase perception capabilities of artificial systems and have the potential to make significant impact in our future economy and society.