Vision-Based Analysis of Crowded Scenes

22 Apr
Friday, 04/22/2011 11:00am to 12:30pm

Robert T. Collins
Pennsylvania State University
Computer Science & Engineering Department

Computer Science Building, Room 151

Detecting and tracking people in a crowded scene is challenging for human observers due to the large number of constantly moving individuals. Over the past five years, we have been developing computer vision algorithms for helping to automate the analysis of crowds. Given video from a single camera view, we have developed a Bayesian approach for simultaneously estimating the number of people in the scene and their spatial locations by sampling from a posterior distribution over crowd configurations. Although naive extension of this approach to multiple views leads to an inefficient sampler, we recently presented a new set of sampling proposals that leverage multiview geometry to more efficiently explore modes of the posterior distribution. In joint work with the Sociology department, we have developed an automated method for discovering which people in a scene are traveling together, and have shown that the results agree well with human perception of pedestrian groups. Finally, we have been developing visualization tools to aid the empirical study of social behavior in crowds.