Crowd counting and density estimation method based on multi-column CNN and adaptive projections onto convex sets

Abstract

Aiming at the low accuracy of crowd counting and density estimation algorithm in the environment of dense crowd with poor characteristics and complex background, a Crowd counting and density estimation method based on Multi-column CNN and adaptive projections onto convex sets was proposed. Firstly, a series of images translated from a single image by a specific value are put into the Multi-column CNN to obtain initial density estimation maps with information difference by sub-pixel displacement. Secondly, the crowd density recognition process is regarded as the sampling and mapping form the real space to density map, and the recognition ambiguity is simulated by the point spread function. In the training of convolutional neural network, the adaptive parameters of projections onto convex sets are trained synchronously to optimize the constraints of convex sets and the degree of image fusion correction. Finally, several density estimation maps with information difference by sub-pixel displacement are fused by adaptive projections onto convex sets to obtain the final density estimation map. The accuracy and robustness of the proposed model are proved by experiments