Fast Crowd Density Estimation in Surveillance Videos without Training

Abstract

Crowd analytics is becoming a highly desirable feature of Intelligent Video Surveillance (IVS) applications. In this paper we propose a new, practical approach that adds very little computational and configuration overhead to an IVS system. The approach extends a standard IVS system, using available video content analysis data and camera calibration information to provide accurate human count estimation in crowded scenarios. The algorithm is viewpoint independent and requires no training for different camera views. The primary output of the algorithm is a real-time crowd density measurement at each image location. This can be further used to detect various crowd related events. Extensive experiments show that the approach is robust and it has been integrated into a commercially available IVS system.