A low latency behavior analysis of human crowds is important for the prevention of crowd disasters. It is critical, in particular, in a security context because of the opportunity to cause large damage through malicious actions as we have seen during the Boston Marathon on April 15, 2013. Ensuring safety through continuous crowd monitoring by humans is impractical in heavily utilized pedestrian scenes and a computer vision-based tracking of each individual in a large crowd is virtually impossible due to the combination of sheer numbers, the proximity between people, and partial occlusion.
Here we consider a method for real time identification of motion anomalies in large, dense crowds. Our method addresses the prohibitive complexity of individual-based motion analysis methods within a vision of a layered, multi-scale framework. Specifically, here we focus on the first stage where coarse grain / macroscopic level analysis identifies and localizes suspected anomalies, allowing the subsequent focus of sensing and computational resources and enabling the use of finer grain methods, based on individual motion analysis, in small suspect areas.
To make real time, coarse grain motion analysis of a large and dense crow motion computationally feasible we make a continuum assumption and model the entire crowd as a continuous fluid. By this approach, crowd dynamics are described by a set of parameterized partial differential equations (PDEs) inspired by the compressible Navier-Stokes equations and include terms for driving purposeful motion, density aversion, multiple desired paths, crowd viscosity effects, and a long observed natural stochasticity, such as in pedestrians collusion avoidance maneuvers. Using numerical simulations we demonstrate that, indeed, this model is capable of reproducing some well-known self-organization phenomena of human crowds.
The sought complexity reduction is achieved by characterized localized crowd behavior in terms of the values and temporal variations of the identified local parameters of governing PDEs. We developed an auxiliary, low order dynamical model for the evolution of those parameters and use that model for the identification of abrupt parameter changes which indicate a local crowd motion anomaly.
Algorithms employing video data, such as our crowd anomaly detection, require a comprehensive video footage database with accompanying high-accuracy labels for tuning and performance testing. Since the accurate labeling of videos is a highly cost intensive process, especially for large crowds, the development of sophisticated algorithms is typically limited to an economized label database to meet project budget constraints. To overcome this limitation, we developed methods to accelerate the generation of high quality video labels through the deployment of a flexible annotation framework with computer-aided labeling support.
Advisor: Professor Gilead Tadmor
Professor Octavia Camps
Professor Mario Sznaier