The Research Of Crowd Density Estimation In Intelligent Surveillance System

With the development of economic society, more and more people tend to move to big cities and the density of population become larger. Recently, stampede event has been reported frequently. It aroused people’s wide concern, especially in the circumstance of religious rites, sports events, celebration parties and other large-scale activities. Therefore, detecting crowd density in public is of great significant. The intelligent video supervise system analyzes crowd density. And according to the analyzing results, security agencies deploy security manpower to prevent stampede. So how to use intelligent video supervise system to analyze crowd density becomes very important. In addition, crowd density estimation has guiding significance in many kinds of tasks like traffic scheduling, market research, building design and so on. Nowadays, using image processing techniques and computer vision techniques to solve the problem of crowd density estimation becomes a research hotspot.The paper focused on crowd density estimation algorithm in intelligent monitoring system. It proposed an algorithm of crowd density estimation by combining pixel feature and texture feature in this paper. The algorithm solved the problem of low precision both in estimating high density crowds when using the method based on pixel feature and in estimating low density crowds when using the method based on texture feature.Firstly, it obtained moving foreground by using an improved background subtraction algorithm. The algorithm made good use of image information in each channel, so that it can be better applied to color videos. Compared to previous algorithms, the method improved detecting precision. It reduced the probability of incomplete object detecting and ensured the integrity of foreground objects. Finally, it divided image frames into low density frames and high density frames according to the proportion of foreground.In the situation of low density, due to the effect of perspective in surveillance videos, the same group of people in different areas has different sizes.In order to avoid perspective effect, it used object contours to replace full objects. Then, it proposed to segment the image into blocks and gave each block a weight to improve the precision of the algorithm. Finally, it divided groups of people into low density situation and the lowest density situation according to the proportion of contour pixels.In the situation of high density, it used gray level co-occurrence matrix to analyze the texture of the image. To obtain gray level co-occurrence matrix, it firstly segmented the image into blocks, and then computed the texture feature of each image. In this way, it reduced the effect of perspective. By experiment, it learnt optimal parameters of gray level co-occurrence matrix and then obtained the texture feature, and it enhanced the precision of detection and described the feature better. Finally, it used support vector machine method to train texture feature and then further divided high density groups of people into medium density, high density and the highest density circumstances.Finally, experiment results demonstrate the method is robust and has higher precision in crowd density estimation.