Multi-image Crowd Counting Using Multi-column Convolutional Neural Network

Abstract

Crowd density estimation is an important task for security applications. It is a regression problem consisting of feature extraction and estimation steps. In this study, we propose to use a modified version of previously introduced multi-column convolutional neural network (MCNN) approach for estimating crowd density. While in the original MCNN approach the same input image is applied to the each column of the network, we first propose to apply a different version of the same input image to extract a different mapping from each column. Second, original MCNN first generates an estimated density map and then performs crowd counting. Therefore, we adopt it for crowd counting and compare its performance with the proposed method. Regression task is performed by support vector regression (SVR) using feature vectors obtained from MCCNN. 2000 images selected from UCSD pedestrian dataset are used in the experiments. The regions of interest (ROI) are filtered out and the pixel values at the remaining regions are set to zero. In order to prevent distortion caused by camera position, perspective normalization has been applied as a pre-processing step which dramatically improves the performance.