Research On Crowd Density Estimation For Scenic Spots Based On Videos

Nowadays,with the development of computing ability and machine learning,the concept of intelligence has penetrated into every corner of society in the field of computer.The concept of intelligent scenic spot also emerged which aims to replace traditional management mode relied on human resources totally,and helps people to manage scenic spots scientifically.Especially in the field of scenic spot surveillance,the traditional model of monitoring is prone to judge mistakenly,not in time and so on,and it's badly in need of computer to help people complete these judgements.The crowd density estimation is one of the most important ingredients of crowd surveillance,and the research on it has lasted for many years.The paper introduces the research situation of this problem in recent years,and summarize the general framework of crowd density estimation.By summarizing the advantages and disadvantages of existing methods and the characteristics of scenic spots,a crowd density estimation method based on multi-feature ensemble learning is proposed.First,this paper researches on the traditional solutions of perspective projection effect,which exist various problems,such as the requirement of field camera calibration,the difficulty of implement on certain features,and the cross-camera capacity.Aiming at these problems,the paper proposes an area-based blocking method,by introducing the scale factor of "area" and combining the weight map method and the blocking method.The different scenes can be divided into the sub image blocks of equal area easily,and the algorithms or models can be transplanted simply.Then this paper evaluates the traditional methods of local estimation,which cannot achieve high accuracy in practical.Because the foreground extraction is not good for complex scenarios,and the single feature is insufficient for the crowd expression.Therefore,this paper proposed a multi-feature ensemble learning method based on a learning strategy,which extracts different features from original images,and learns the outputs of base learners to improve the accuracy of regression.Then this paper analyzes the traditional summation method that estimates the global count from local counts.This method lead to another estimation errors easily.This paper designs a horizontal overlapping block method,which combines support vector machine.It effectively solves this error caused by the traditional direct summation method by the learning strategy.Finally,in the experiment,each elements of the system are tested by different methods,which explains the reasons for choosing these methods.Then the proposed method is compared with other methods in the experimented scenic spot datasets and a public dataset,and shows better accuracy.At last,the proposed method is tested on other scenes of the experimented scenic spot,and achieves more than 90%of the classification accuracy using the same model,which indicates the good expansibility of the proposed method.