Saliency Based Dynamic Scene Analysis

Abstract

In present era of continuously increasing video data, it is needed to abandon manual analysis of dynamic scenes. To handle this large amount of data, some type of automation has to be used, but this is highly complex task. Our brain handle such huge amount of data by allocating most of it part to this task and this task become slightly lighter due to anatomy of our eye. We shift our gaze to the specific location in image to dedicate more computational resource, rather than to consider whole image. To replicate the behavior of human attentional system various models have been proposed. Most of the models are based on low-level features and some are based on combination of both high and low-level features. These models focus on static images. And those static models don’t behave well enough in accordance with human gaze movement for dynamic scene. There are significant relation between salient region and other regions of the image which can be affected spatially and temporally, such dynamic data is significantly complex and we cannot depend only on motion energy alone. Modelling these relations and keep model bio-plausible are the challenges addressed by this thesis. We here investigate that a blended model of dynamic features, high and low level features outperforms low-level and other dynamic models in guessing human fixation on dynamic visual input, on the basis of movement of human eye recordings while observing videos of natural outdoor scenes, most of which contained at least fifteen persons.