National Technical University of Athens (NTUA)
Καινοτόμες τεχνικές όρασης υπολογιστών και μηχανικής μάθησης για αναγνώριση συμπεριφοράς από ακολουθίες βίντεο και προσαρμογή τους σε π...
Abstract
dc:descriptionThis thesis aims at proposing novel computer vision and machine learning techniques for human behavior recognition from video, emphasizing on activities forming workflows. The problem involves significant challenges, such as occlusions, frequent illumination changes and outliers, whereas the visually complex environment of our use case induces additional difficulties. Typical object based methods that rely on detection or tracking tend to fail because of the high complexity of the observed scenes. In order to bypass these error-prone stages, we propose the extraction of holistic features directly on the image level for scene representation. Furthermore, we examine the applicability of fused schemes of the employed classifiers (hidden Markov models) in an endeavor to exploit redundancies from multiple camera streams so as to solve occlusions, whereas we also scrutinize the effectiveness of the multivariate Student-t distribution as observation likelihood (instead of the Gaussian) for higher tolerance to outliers. In the sequel, a series of new techniques for online behavior recognition are proposed. The first one is based on sequence segmentation, classification of the segments through HMM classifiers and incorporation of a priori knowledge via a genetic algorithm (GA-HMM). The second method is based on a combination of bayesian filtering and HMM and does not require a separate segmentation algorithm, while the third one focuses on concurrent activity recognition following a top-down event-driven Region of Interest based approach. Another contribution of this thesis lies in the introduction of the concept of Evaluative Rectification, which exploits an expert user’s feedback regarding the correctness of the classification results of the employed recognition methods and follows a feedforward and adaptive neural network based approach in order to improve future results in the direction of decreasing the overall classification error. Finally, we examine the possibility of exploiting the benefits involved in cloud computing and propose a cloud platform endowed with modern workflow management mechanisms for an effective and efficient application of the proposed methods in real-world large-scale installations, where online and real-time requirements are posed.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Βουλόδημος, Αθανάσιος
- Voulodimos, Athanasios
Subjects
dc:subject × 24- Αναγνώριση συμπεριφοράς
- Ολιστικά χαρακτηριστικά
- Κρυμμένα μαρκοβιανά μοντέλα
- Πολυρευματική σύμμιξη
- Ευρωστία σε έκτοπα
- Συνόψιση βίντεο
- Υπολογιστική νέφους
- Γενετικοί αλγόριθμοι
- Behavior recognition
- Holistic features
- Hidden Markov models
- Multistream fusion
- Robustness to outliers
- Video summarization
- Cloud computing
- Genetic algorithms
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Natural Sciences
- Computer and Information Sciences
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/26947