CrowdVAS-Net: A Deep-CNN Based Framework to Detect Abnormal Crowd-Motion Behavior in Videos for Predicting Crowd Disaster
Tanu Gupta, Vimala Nunavath, Sudip Kumar Roy · 2019
With the increased occurrences of crowd disasters like human stampedes, crowd management and their safety during mass gathering events like concerts, congregation or political rally, etc., are vital tasks for the security personnel. In this paper, we propose a framework named as CrowdVAS-Net for crowd-motion analysis that considers velocity, acceleration and saliency features in the video frames of a moving crowd. CrowdVAS-Net relies on a deep convolutional neural network (DCNN) for extracting motion and appearance feature representations from the video frames that help us in classifying the crowd-motion behavior as abnormal or normal from a short video clip. These feature representations are then trained with a random forest classifier. Furthermore, a dataset having 704 video clips having dense crowded scenes have been created for performance evaluation of the proposed method. Simulation results confirm that the proposed CrowdVAS-Net achieves the classification accuracy of 77.8% outperforming the state-of-the-art machine learning models. Moreover, this framework can reduce the video processing and analyzing time up to 96.8% compared to state-of-the-art techniques on the larger dataset. Based on our results, we believe that our work will help security personnel and crowd managers in ensuring the public safety during mass gatherings with better accuracy.