Collectiveness analysis with visual attributes

Abstract

Individuals within crowd scenes tend to move unconsciously in a collective manner. The nature of this phenomenon surges motivations in collectiveness analysis to quantify and detect the collective behavior within the crowds. Due to the complexity of crowd scenes, most studies focus on the collective motion of individuals. However, it requires the extraction of temporal information, i.e., motion attributes, in consecutive video frames. Based on the approach, collectiveness analysis relies on the total number of motion attributes that could not represent the total number of individuals and limits mid-level understanding within crowd scenes. Alternatively, this study proposes a novel framework for collectiveness analysis using visual attributes. It is based on visual attributes extraction approach to facilitate individual-level understanding based on still image input. By localizing individuals and classifying individuals’ head pose, the proposed framework alleviates the need to rely on temporal information and explores topological relationship propagation among individuals to infer collectiveness analysis. Inclusive experiments on various crowd densities illustrate the aims of the proposed framework to infer high-level crowd analysis with visual attributes. Its efficacy is evaluated on real crowd scenes and compared with the state-of-the-art approaches including achieving estimation of group with Average Difference (AD) = 1.68 and Mean Square Error (MSE) = 1.71. Its potential applicability is demonstrated in the context of crowd estimation and collectiveness analysis.