Marked point processes for crowd counting

Abstract

A Bayesian marked point process (MPP) model is developed to detect and count people in crowded scenes. The model couples a spatial stochastic process governing number and placement of individuals with a conditional mark process for selecting body shape. We automatically learn the mark (shape) process from training video by estimating a mixture of Bernoulli shape prototypes along with an extrinsic shape distribution describing the orientation and scaling of these shapes for any given image location. The reversible jump Markov Chain Monte Carlo framework is used to efficiently search for the maximum a posteriori configuration of shapes, leading to an estimate of the count, location and pose of each person in the scene. Quantitative results of crowd counting are presented for two publicly available datasets with known ground truth.

Crowd Detection with a Multiview Sampler

“…Among monocular approaches for pedestrian detection [4][5][6][7][8][9], classifier-based methods are very popular [7][8][9] and sampling-based methods have also been shown effective for crowd detection [2,3,10] as well as generic object detection [11,12]. Within the sampling framework, various efficient, data-driven sampling strategies have been proposed.…”

“…Within the sampling framework, various efficient, data-driven sampling strategies have been proposed. For example, Zhao and Nevatia [2] use a head detector to guide location estimates and Ge and Collins [3] learn sequencespecific shape templates to provide a better fit to foreground blobs. We extend the sampling framework to a unified approach that can detect people visible in a single view or in multiple views.…”

“…Below we describe a baseline multiview sampler directly extended from local birth, death, and update proposals commonly used in single view samplers [2,3].…”

“…Our approach is motivated by the success of previous generative models for people detection [1][2][3]. Due to the great flexibility offered by sampling-based inference methods, our crowd model can accommodate inter-person dependencies that otherwise would be intractable to infer because of their inherent combinatorics.…”

“…Efficient sampling strategies are the key to performance in practice. Although various data-driven proposals have been designed in the single view context to guide hypothesis generation [2,3], to our knowledge we are the first to explore multiview geometry constraints for efficient sampling.…”