Using facial expression recognition for crowd monitoring.

Abstract

In recent years, Crowd Monitoring techniques have attracted emerging interest in the eld of computer vision due to their ability to monitor groups of people in crowded areas, where conventional image processing methods would not suffice. Existing Crowd Monitoring techniques focus heavily on analyzing a crowd as a single entity, usually in terms of their density and movement pattern. While these techniques are well suited for the task of identifying dangerous and emergency situations, such as a large group of people exiting a building at once, they are very limited when it comes to identifying emotion within a crowd. By isolating different types of emotion within a crowd, we aim to predict the mood of a crowd even in scenes of non-panic. In this work, we propose a novel Crowd Monitoring system based on estimating crowd emotion using Facial Expression Recognition (FER). In the past decade, both FER and activity recognition have been proposed for human emotion detection. However, facial expression is arguably more descriptive when identifying emotion and is less likely to be obscured in crowded environments compared to body pos- ture. Given a crowd image, the popular Viola and Jones face detection algorithm is used to detect and extract unobscured faces from individuals in the crowd. A ro- bust and efficient appearance based method of FER, such as Gradient Local Ternary Pattern (GLTP), is used together with a machine learning algorithm, Support Vec- tor Machine (SVM), to extract and classify each facial expression as one of seven universally accepted emotions (joy, surprise, anger, fear, disgust, sadness or neutral emotion). Crowd emotion is estimated by isolating groups of similar emotion based on their relative size and weighting. To validate the effectiveness of the proposed system, a series of cross-validation tests are performed using a novel Crowd Emotion dataset with known ground-truth emotions. The results show that the system presented is able to accurately and efficiently predict multiple classes of crowd emotion even in non-panic situations where movement and density information may be incomplete. In the future, this type of system can be used for many security applications; such as helping to alert authorities to potentially aggressive crowds of people in real-time.