Monitoring Crowding Level: Visual Learning in Virtual Environments

The paper presents a solution to the problem of crowding estimation in large environments using multiple visual sensors. The proposed approach presents innovations addressing several information processing stages of a surveillance system. A novel, robust way to manage shadows in low resolution, color or monochrome images, is introduced providing reliable background subtraction. The resulting robust foreground pixel count is mapped into people count by a new learning algorithm that samples the environment in an adaptive way by directing a human probe to stabilize its results. The maps from multiple sensors are then automatically aligned into a global, two dimensional, reference coordinate system without requiring explicit sensor calibration. This feature supports information integration and the development of fault tolerant surveillance systems. Algorithm development and performance assessment rely on an innovative methodology based on photo-realistic synthetic museum environments populated by virtual visitors. The simulator provides a flexible test bed for the investigation of automatic surveillance systems. As a sample application, the resulting occupancy maps are then used to characterize people flow patterns at different crowding levels exposing possible visiting bottlenecks.