Conditional marked point process-based crowd counting in sparsely and moderately crowded scenes

Crowd density estimation for counting persons, or for determining interactions among persons, groups of people, or crowds has been a challenging problem since persons can be occluded by other persons in (highly) crowded situations. The successful development of such techniques has diverse purposes, such as reassigning limited resources (e.g., public transportation) properly by counting floating population or categorizing the type of events based on the identification of crowd interactions. While existing counting approaches are mostly based on regression models that directly map features to the corresponding class labels, we propose a conditional marked point process (CMPP)-based approach to count individual persons even in moderately crowded scenes. We use a mixture of Bernoulli shape, which is a stochastic model, estimated from the training set with extrinsic shape distribution that determines the size of a shape for the given location in an input image to count the proper number of persons in different types of scenes. The experiment was carried out on PETS2009 which is a well-known public dataset. It was concluded from the experimental results that the proposed approach can be an alternative to the conventional MPP-based approaches.