Um mecanismo de atenção visual integrando evidências espaciais e temporais.

Abstract

Biological vision systems have mechanisms that focus on the extraction of the most relevant information for performing a given visual task, so that the overall computational effort is reduced. Inspiredby these mechanisms, Visual Attention emerges as the area of Computer Vision that is mainly concerned with the processing of visual scenes, searching for the most salient regions (which are the mostimportantto be analyzed). Within this context, the present work proposes a new Visual Attention model, which integrates different approaches: Spatial (or Static)bottom-up Attention, Temporal (or Dynamic) Attention and Stereo Vision. This work also develops, for the first two approaches, an implementation that is validated through a series of experiments.Furthermore, this work presents a strategy for the segmentation of moving objects in visual scenes, as part of the proposed model, and a casestudy, involving the utilization of temporal evidences from the developed Visual Attention system in the problem of detecting sharp transitions in video sequences. The results have shown that the strategy proposed for the temporal feature extraction and for the detection of moving objects was a simple and versatileway to perform motion detection and segmentation in videos. With regards to the experiments involving the detection of sharp video transitions, a performance evaluation revealed low error rates. Finally, the integration of spatial features into the above context yielded an interesting approach to combine evidence from both Spatial and Temporal Visual Attention.