Region-Oriented Visual Attention Framework for Activity Detection

摘要

This paper proposes a framework, based on a spatio-temporal attentive mechanism, for automatic region-of-interest determination, corresponding to events in video sequences of natural scenes of dynamic environments. We view this work as a preliminary step towards the solution of high-level semantic event analysis. More specifically, we wish to detect a visual event within a cluttered scene, without intensive training algorithms. In contrast to event detection methods used in the literature, which drive attention based on motion and spatial location hypothesis, in our approach the visual attention is region-driven as well as feature-driven. For this purpose, a two stages attention mechanism is proposed. In a first phase, spatio-temporal activity analysis extracts key-frames from the image sequence and selects salient areas within these frames. The three types of visual attention features are used, namely, intensity, color and motion. Consequently, the selected areas are further processed to determine the most active region, based on a newly defined region saliency measure. Qualitative and quantitative results, using the proposed framework, are illustrated envisaging the application domain of change detection in automated visual surveillance.