Counting People in Crowds with a Real-Time Network of Simple Image Sensors

Estimating the number of people in a crowded environment is acentral task in civilian surveillance. Most vision-based countingtechniques depend on detecting individuals in order to count, anunrealistic proposition in crowded settings. We propose analternative approach that directly estimates the number of people.In our system, groups of image sensors segment foreground objectsfrom the background, aggregate the resulting silhouettes over anetwork, and compute a planar projection of the scene's visualhull. We introduce a geometric algorithm that calculates bounds onthe number of persons in each region of the projection, afterphantom regions have been eliminated. The computationalrequirements scale well with the number of sensors and the numberof people, and only limited amounts of data are transmitted overthe network. Because of these properties, our system runs inreal-time and can be deployed as an untethered wireless sensornetwork. We describe the major components of our system, and reportpreliminary experiments with our first prototype implementation.