Crowd density estimation using multiple features categories and multiple regression models

Abstract

This paper proposes a new single camera automatic crowd density estimation method for overcoming the linearity problem and enhancing the counting prediction accuracy. For overcoming the linearity problem, a combination of features including segmented regions properties, texture, edge, and SIFT keypoints are extracted from the segmented foreground regions. These features are normalized to solve the perspective distortion problem. The complete feature set and a partial of this set are utilized to train multiple regression models to predict the crowd density per region and enhance the prediction accuracy. Moreover, the cross-validation technique is used to select the training and testing datasets in order to increase the prediction accuracy. Categories of features are ranked based on a set of metrics reflecting their robustness with each regression model used. Using the ground truth data provided by the UCSD crowd dataset, the experimental results show that the proposed crowd density estimation method are more robust and provide crowd counting prediction with higher accuracy than previous methods.