A Method For Crowd Counting Based On Multi-Classifier Ensemble

Nowadays,with the rapid economic development,urbanization process continues to advance,the number of urban population has also increased dramatically.In particular,a country with large population like China,a number of second tier cities,a series of high-density population problems have become increasingly prominent.Crowded places,such as large shopping malls,outdoor plazas,waiting halls and so on,crowd congestion occurs frequently.The estimation of population size and population density in public places has important social significance and application prospect.The traditional method of population statistics is artificial statistics,that is,through the human eye to identify and count.The disadvantages of this method are obvious : Manual statistical method will consume a lot of manpower,material resources,financial resources,higher costs,resulting in great waste of resources;When the staff is in a state of fatigue,the accuracy of the statistical results will be significantly affected.While with statistical algorithms of the number of people based on intelligent video surveillance,the above problems can be solved well,and only need to use the existing camera,without the addition of a large number of additional equipment,low cost,good real-time,easy to use.This paper proposed a new crowd counting algorithm based on the multi-classifier ensemble strategy.Firstly,the moving foreground of pedestrians is extracted,the foreground area after correction of the perspective effect is calculated,and the effective Harris corner and SURF point information are obtained to calculate the occlusion coefficient which reflects the occlusion degree of the pedestrians.Therefore,this algorithm can construct the feature vectors and set up a regression model by using them as the input variables of Back Propagation(BP)neural network.By extracting the HOG feature of pedestrians,the corresponding pedestrian detector is trained by the Adaboost cascade classifier to detect the pedestrians and count the number in per frame.Finally,based on the first two classifiers,the combination classifier is constructed by the stacking strategy with adaptive adjustment of weights.The experimental results show that the proposed algorithm is superior to other existing algorithms,with a strong adaptability for complex scenes and a good real-time performance.