A biologically plausible system for detecting saliency in video

Show full item record

Title: A biologically plausible system for detecting saliency in video
Author: Burlone, David
Abstract: Neuroscientists and cognitive scientists credit the dorsal and ventral pathways for the capability of detecting both still salient and motion salient objects. In this work, a framework is developed to explore potential models of still and motion saliency and is an extension of the original VENUS system. The early visual pathway is modeled by using Independent Component Analysis to learn a set of Gabor-like receptive fields similar to those found in the mammalian visual pathway. These spatial receptive fields form a set of 2D basis feature matrices, which are used to decompose complex visual stimuli into their spatial components. A still saliency map is formed by combining the outputs of convoluting the learned spatial receptive fields with the input stimuli. The dorsal pathway is primarily focused on motion-based information. In this framework, the model uses simple motion segmentation and tracking algorithms to create a statistical model of the motion and color-related information in video streams. A key feature of the human visual system is the ability to detect novelty. This framework uses a set of Gaussian distributions to model color and motion. When a unique event is detected, Gaussian distributions are created and the event is declared novel. The next time a similar event is detected the framework is able to determine that the event is not novel based on the previously created distributions. A forgetting term is also included that allows events that have not been detected for a long period of time to be forgotten.
Record URI: http://hdl.handle.net/1850/1923
Date: 2006

Files in this item

Files Size Format View
DBurloneThesis052006.pdf 4.368Mb PDF View/Open

The following license files are associated with this item:

This item appears in the following Collection(s)

Show full item record

Search RIT DML


Advanced Search

Browse