Crowd Density Estimation Based On Texture Analysis

With the rapid development of the economy and society, varieties of large squares and entertainment facilities emerge in an endless stream, and crowd activities of large scale have become common. At the same time, the disasters caused by high density-crowd happened frequently. In such a situation, making a crowd density estimation of the monitor area and taking evacuation measures to high density area become important.The most commonly used method of crowd density estimation is to extract features of crowd image firstly, and then to classify the features using a classifier to get the level of crowd density. Texture features extracted from gray level co-occurrence matrix(GLCM) is the traditionally used features, but these kinds of texture features only use the gray information of the crowd image. The correct classification rate is not high. In view of this situation, this paper proposes a method combining the gray level co-occurrence matrix’s(GLCM) feature parameters and gradient-gray co-occurrence matrix’s(GGCM) feature parameters, constituting a feature vector of twenty dimensions. In this way, not only gray information of the crowd image, but also the gradient information is used, which get a better classification result.In the stage of classifying the features, the traditional approach is using the one against one SVM, one against more SVM or directed acyclic graph SVM. But this method requires many sub SVMs and the time used in training and testing is long. To solve this problem, this paper proposes an improved SVM-the shortest distance-based binary tree SVM to reduce the number of sub SVM and shorten the time.In addition to the SVM, the paper proposes using extreme learning machine (ELM) to classify the texture features. ELM is a new algorithm based on single hidden layer neural network model in recent years replacing the traditional gradient descent method. ELM algorithm assigns the value of input weights and bias randomly, and then calculate the output weight. Experiments show that the ELM can achieve better results in real-time.Moreover, the theory of the SVM and ELM is compared in this paper and the similarities and differences of these two classifiers are analyzed. Experimental results demonstrate the accuracy of these two kinds of classifiers is quite, but the real-time performance of ELM is better, whose training speed is about ten times that of the SVM and test speed is about five times that of the SVM.