An Intelligent Extraversion Analysis Scheme from Crowd Trajectories for Surveillance
In recent years, crowd analysis is important for applications such as smart cities, intelligent transportation system, customer behavior prediction, and visual surveillance. Understanding the characteristics of the individual motion in a crowd can be beneficial for social event detection and abnormal detection, but it has rarely been studied. In this paper, we focus on the extraversion measure of individual motions in crowds based on trajectory data. Extraversion is one of typical personalities that is often observed in human crowd behaviors and it can reflect not only the characteristics of the individual motion, but also the that of the holistic crowd motions. To our best knowledge, this is the first attempt to analyze individual extraversion of crowd motions based on trajectories. To accomplish this, we first present a effective composite motion descriptor, which integrates the basic individual motion information and social metrics, to describe the extraversion of each individual in a crowd. The social metrics consider both the neighboring distribution and their interaction pattern. Since our major goal is to learn a universal scoring function that can measure the degrees of extraversion across varied crowd scenes, we incorporate and adapt the active learning technique to the relative attribute approach. Specifically, we assume the social groups in any crowds contain individuals with the similar degree of extraversion. Based on such assumption, we significantly reduce the computation cost by clustering and ranking the trajectories actively. Finally, we demonstrate the performance of our proposed method by measuring the degree of extraversion for real individual trajectories in crowds and analyzing crowd scenes from a real-world dataset.