Large Crowd Count Based On Improved SURF Algorithm

With the world population increasing, the large-scale group activities are more and more. Because the population is too dense and the accidents are caused constantly. So the statistics on the number of monitoring system is very necessary. When the video number reaches a certain range, it can play a role in the early warning to avoid the occurrence of accidents. Meanwhile, the manpower and material resources get reasonable arrangement and the resources get a unified deployment. But the traditional statistics can not very good solve the effect mass caused by the occlusion and projective deformity, so this paper puts forward to the number of video statistics based on an improved SURF feature extraction algorithm.In the process of number statistics, firstly this paper extracts the foreground image after binarization. Secondly the multi feature of the foreground image are extracted, and then to carry the perspective rectification, so as to form a feature vector. Finally the feature vector is regression prediction. In the foreground image extraction process, the video image is in pretreatment. To get the background modeling is used the average model method. The foreground image is extracted by difference methods. Because the light will appear under the shadow, so it has to have a shadow suppression of foreground image. Then the foreground image is used binary morphological image processing methods. In the process of feature extraction, in order to solve the serious occlusion problems existing in the high density population count, so the image gray level co-occurrence matrix feature is extracted. And due to the problems of the camera projective distortion, it is introduced the linear interpolation weight correction parameters to perspective correction of SURF features and pixels. A feature vector is constructed for the extracted multiple eigenvalues. Finally, the optimized feature vector is obtained by support vector regression to get the regression template, which can predict the number.Experiments show that the improved feature value can represent the number of people. And compared with the traditional people counting algorithms, the detection accuracy of the method is improved. The mean relative error is less than 15%, which can effectively solve the occlusion and projective deformity these problems. It can accurately estimate the number of people, and is more suitable for the large-scale population count.