Research Of Crowd Density Estimation In Video Surveillance Scene

Recent years, with the acceleration of global urbanization, urban crowds density is becoming bigger and bigger, crowds gather more and more, dense crowds moving would easily cause serious casualty accident such as congestion, stampede. Such disaster occurred frequently in the world. Therefore, estimation on crowd density in video surveillance scene is of profound significance.The density estimation method based on pixel statistics is relatively simple and the algorithm complexity is low. However, when the crowd density is high, this algorithm is not suitable. The method based on texture feature analysis can make full use of the texture information. In order to solve the problem above, in this paper, an improved crowd density estimation algorithm is proposed. The threshold is used to divide the images into low density crowds and medium and high density crowds, and uses two different crowd density estimation methods.For low crowd density scenes, this paper adopts the method based on the foreground pixels and the linear fitting method. First of all, the weighted average method is adopted to gray the crowd images, and then the median filtering method is used to eliminate noise and outliers, finally, the background images are constructed, and the preliminary crowd foreground images are obtained by background subtraction operation, and the final crowd foreground images are obtained by morphological operation. Least square linear fitting is adopted to estimate the linear relationship between the number of foreground pixels and the number of people.For median and high crowd density scenes, this paper uses the crowd density estimation method based on texture feature analysis and support vector machine (SVM). The texture features of the images are extracted by gray level co-occurrence matrix, and the optimal texture features of the images are selected through the experimental study to the gray level co-occurrence matrix structural parameters, energy, contrast, entropy and correlation are selected. For the classification of pattern recognition, SVM is adopted to classify the median and high density crowds, and the SVM model is set up by using the training sample according to the classification rules, the optimal penalty parameter C and the kernel function are selected through experimental study. Then SVM is used to classify and obtain the classification results, and the crowd density estimation in video surveillance scene is completed.Finally, in order to verify the feasibility and effectiveness of the proposed algorithm, the crowd video was carried out, and the low density crowd estimation was completed through the least square linear fitting. The classification accuracy of the median and high density crowd reaches 90%. The experimental results show that this method is effective, and it can provide powerful help for the protection of public safety.