Research On The Number Of People In Public Places Based On Video

In recent years,with the ever fast development of science and technology,continuous social progress,economic growth and the continuous improvement of image and video processing algorithms,there are increasingly more extensive applications of various images and videos in real life.In the field of image processing,it has become a hotspot to study people counting in public locations based on video,which has attracted broad attention.The people counting work has potential value both in business and in real life.Based on that,the subject of people counting in public locations is studied in this paper based on video technology.First of all,this paper summarizes related works of people counting methods,and generally divides the people counting methods into two categories of individual statistics and group feature statistics for overview.Then,reasonable division of people flow density in public place is conducted for development of subsequent works;next,for the people counting work in public place,this paper innovatively classifies the target scenarios into the two major types of sparse scenario and dense scenario for research,and for the first type,the individual feature method is adopted for statistic research.In the sparse scenario,by using individual features for statistics,it can not only improve the statistic accuracy,but also benefit subsequent application extension based on the result of this research,such as using individual features for target tracing and face recognition;for people counting under dense scenario,considering concentrated flow of people,the recognition accuracy and timeliness have not been ideal so far,so for now,A relatively effective solution is to conduct statistics based on group features.The main work of this paper include:1.The NSCT and CS-LBP feature-based algorithms are integrated for the first time,which are combined with the human head profile area localization to be used for people counting under sparse scenario.First of all,the general characteristics of profile ellipse are used for fast localization and extraction of human head,and for the target area,the NSCT feature-based algorithm is used for filter decomposition of image in different scales and directions;secondly,CS-LBP can provide good robustness for illumination,extract feature histogram of the blocks of high-and low-frequency subband information after decomposition,and conduct cascade connection to obtain a complete vector histogram of human head features.For the high dimension problem of image features after processing,the principal component analysis(PCA)method is adopted for dimension reduction processing,which can finally effectively improve the real-time performance of recognition;finally,the traditional support vector machine(SVM)method is used for people counting.2.Among researches on people counting in dense public locations,in order to effectively increase the detection efficiency and satisfy the requirement for practical application,the improved S-Harris corner detection algorithm is proposed.In the meantime,based on the idea of Kalman filtering,the adaptive dynamic model regression method is adopted for the statistic work.First of all,according to the insufficiencies of Harris corner algorithm in people counting,an adaptive gray difference idea is proposed,and the concept of integral image is introduced to solve the defects in noise resistance and real-time performance of operation.Secondly,considering the first-order static model tends to generate high error during the people counting process,this paper proposes a dynamic linear model regression method for people counting.This method believes that as time goes by,there is certain coefficient relation in the proportional relation between the corner number and the number of people in each frame,and this coefficient has certain correlation with the corner information between neighboring video frames.In the meantime,in order to eliminate redundant corners generated during the corner statistics process,this paper adopts the frame difference method to filter out the static corners;finally,the dynamic first-order linear model regression is used to calculate the number of people.