Fuzzy qualitative approach to address uncertainty in human motion analysis / Lim Chern Hong

• Main Author: Lim, Chern Hong
• Format: Thesis
• Published: 2015
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://studentsrepo.um.edu.my/5917/1/thesis.pdf; http://studentsrepo.um.edu.my/5917/2/Turnitin_Originality_Report.pdf; http://studentsrepo.um.edu.my/5917/

Human motion analysis is one of the most active researches in computer vision society nowadays due to its wide spectrum of applications. Current researchers have been focused on implementing sophisticated algorithms with the goal to achieve good recognition rate but such work are limited to some constraints or assumptions. As a consequence, these systems are impractical to deploy in real-world environment due to the abounded uncertainties in the human motion analysis pipeline such as human size variation, viewpoint variation, and classification ambiguity. Failing in handling these uncertainties could affect the overall system performance. In this thesis, fuzzy qualitative reasoning is studied to address the above uncertainties. Human modelling is the enabling step in the human motion analysis system where the identified person from a video camera will be projected and represented in a better model to ease the latter processes such as feature extraction. Improper care on the variation of human size and camera positions from the ground might results in a defect human model such as inconsistent human size, and odd human shape. Such defects will hinder the feature extraction process and the error in this step might be cumulated in the rest of the pipeline and deteriorate the overall system performance. In this thesis, fuzzy qualitative Poisson human model is proposed to generalize the human model in terms of sizes and camera viewpoints. Besides that, to recognize an action with independent to the human viewpoint is a great challenge in human motion analysis, but remains unsolved due to its inherent difficulty. Most state-of-the-art methods are found to be impractical where multi camera system is required to serve the purpose. In this context, view specific action recognition framework is proposed to capture and construct the view specific action model for the objective to achieve view invariant human action recognition within single camera. In the framework, a novel human contour namely fuzzy qualitative human contour is proposed for view estimation which helps in the construction of the view specific action model. Action recognition is the final step in the human motion analysis pipeline where the aim is to infer the action or activity from the video. However, classification ambiguity could abounded in this step such as the confusion in viewpoint, action, and scene context due to some similarity factors. These cases are denoted as non-mutually cases in the thesis as their results could not be fully distinguished from the others. Hence, a crisp or binary classifier may not be so effective to deduce the final output for these cases. As a solution, fuzzy qualitative rank classifier is proposed to model the non-mutually exclusive case in the training step and infer with the multi-label and ranking result. This is intuitively reflecting how human decision is made towards the ambiguous case. In addition, dynamic fuzzy qualitative rank classifier is proposed as the extension to overcome the heuristic method in the learning step. In summary, the collective impact of the above contributions will constitute to achieve a more practical and feasible framework towards the human motion analysis applications. Particular video surveillance system that ensure the public safety and lead to a better and safer society.