Transcript Slide 1
Marked Point Processes for Crowd Counting Weina Ge and Robert T. Collins Computer Science and Engineering Department, The Pennsylvania State University, USA We consider a crowd scene as a realization of an MPP MOTIVATION Learning Intrinsic Shape Classes • Model the shapes using a mixture of Bernoulli distributions • Automatically determine the number of mixture components by Bayesian EM • Delineate pedestrians in a fg mask using shape coverings • Bayesian approach Binary mask “Soft” mask MPP prior EM iterations Bayesian EM with Dirichlet prior Combined with likelihood A shape covering • Adapt to different videos by learning the shape models Training samples • Our MPP combines a stochastic point process with a conditional mark process to model prior knowledge on the spatial distribution of an unknown number of pedestrians Automatically learned shapes RESULTS Estimating Extrinsic Parameters π(θi|pi) π(wi, hi|pi) robust regression determine location, scale, orientation Original image determine body shape A rectangular covering Foreground blobs Blob orientation axes of a sequence Inliers found by RANSAC Total # People Detection Rate FP Rate CAVIAR 3728 Soccer 1258 0.92 0.84 0.02 0.06 Blob orientation axes in a frame Vertical vanishing point • Estimating the MPP Configurations by RJMCMC RJMCMC: stochastic mode seeking procedure with two different types of proposals 1. update shape/location local updates to a current configuration 2. birth/death jumps between two configurations of different dimensions CONTRIBUTIONS • An MPP with a conditional mark process to model known correlations between bounding box size/orientation and image location • Bayesian EM for automatic learning of Bernoulli shapes For more info: http://vision.cse.psu.edu/projects/mpp/mpp.html