A Novel Deep Architecture for Multi-Task Crowd Analysis

Abstract

Recently, crowd analysis has become an essential tool for crowd disaster management. The crowd analysis is not a single task but is a collective implication of several related tasks like crowd behavior analysis, crowd counting, and crowd flow analysis. Although different models for individual tasks have been proposed in the literature, there is a lack of a multitasking framework for crowd analysis. One of the main reasons could be the availability of the multitasking crowd analysis dataset. To this end, this paper created a multitask crowd analysis dataset using two publicly available crowd behavior datasets (MED, GTA) and proposed a novel deep architecture for multitasking crowd analysis. Two different crowd analysis tasks are considered, i.e., crowd behavior (normal and panic) and crowd counting. Around 89,000 frames were annotated for obtaining ground-truth crowd counts. In addition to this, a two-stage learning mechanism is proposed. In the first stage, a novel deep model is proposed that extracts high-level spatial-temporal features from the multiscale low-level spatial-temporal features and is used to learn normal crowd behavior and counting. The second stage utilized the features of the deep model and inputted them to the one-class support vector machine (OC-SVM) for crowd panic detection. The obtained results are compared with state-of-the-art and show its effectiveness in multitasking crowd analysis.