Human action recognition using slow feature analysis / Bardia Yousefi

Bardia, Yousefi (2016) Human action recognition using slow feature analysis / Bardia Yousefi. PhD thesis, University of Malaya.

Preview

PDF (Thesis PhD) Download (73Mb) | Preview

Abstract

Studies on computational neuroscience through functional magnetic resonance imaging and following human visual systems state that the mammalian brain pursues two distinct pathways in the model. These pathways are designed to analyze not only motion information (optical flow) but also the ventral processing stream in the brain that proceeds with form features, in which Gabor wavelet is widely used. The original model of the mammalian visual system represents two independent pathways, which become a subject of interest among researchers. Model development is performed via systematic organization, where the active basis model is added into the ventral processing stream. The Gabor wavelet-based and supervised method is efficient in terms of Gabor beam utilization and object recognition-directed task through form pathway. In addition, the motion information that is generated via optical flow in motion pathway is stabilized through applying the fuzzy membership scoring, which delays the changes in optical flow outcomes and provides further robustness to the system. The interaction between these processing pathways is another substantial matter implied in the model. The cross-connection of the two pathways is implied throughout the present research via direct consideration, such as shared sketch algorithm and optical flow information, fuzzy max-product involvement, and scoring among each other. In addition, the model is considered a form information through active basis model based on incremental slow feature analysis (denoted as slow features). In this study, the motion perception in human visual system comprises fast and slow feature interactions, which render biological movement understandable. Primarily, a form feature is defined. This feature biologically follows the visual system through applying active basis model and incremental slow feature analysis for extraction of the slowest form features of human object for ventral stream. The interaction is considered within the time that provides valuable features to recognize biological movements. Incremental slow feature analysis provides a chance for fast action prototypes and bag-of-word techniques, and opens a new perspective to recognize the original biological movement model. Episodic observation is required to extract the slowest features. However, fast features of dorsal processing pathway through episodic ventral analysis update the processing of motion information. Experimental results in the development of the biological movement model indicate promising accuracies for proposed improvements and favorable performance on different datasets (KTH and Weizmann). The results also provide promising direction on this area.

Actions (For repository staff only : Login required)