Research And Implementation Of People Counting And Crowd Density Detection For Intelligent Surveillance

With the rapid development of computer vision technology,video content analysis technology is rapidly emerging in the field of monitoring.Academia and industry have invested tremendous energy and financial resources for the development of related technologies.At the same time,as the process of urbanization accelerates,there are more and more large-scale crowd gathering.On the one hand,people want to obtain valuable information from the crowd to assist in management and decision-making.On the other hand,they also have to face the crowd safety problems caused by overcrowding.Therefore,the human traffic statistics technology and crowd density detection technology have become a hot area of video surveillance,as well as a difficult research.The research of this paper mainly focuses on these two aspects.First of all,this paper reviews a large amount of literature,introduces the status quo of human traffic statistics and crowd density detection technology at home and abroad,and found that most of the current research on traffic statistics algorithms are based on the scene of the vertical camera and close-range,however In reality,many scenes do not meet the camera vertical erection conditions.Therefore,in the aspect of human traffic statistics,this paper focuses on the human traffic statistics method when the camera is not set up vertically and the distance is long.Its main process can be divided into three parts: moving target segmentation,human target detection,multitarget tracking counting.In the aspect of moving object segmentation,this paper studies the background modeling algorithm of Vi Be deeply,simplifies the initialization method of background model under the premise of meeting the application requirements and improves the updating strategy of background model,thus solving the problem of "ghosting" eliminates slow in the original algorithm and make the algorithm more than double the speed of the original algorithm.In the aspect of detection of human head,this paper studied and contrasts the human head detection effect when Adaboost algorithm is combined with multiple features respectively.Finally,based on the experimental results,Adaboost algorithm based on LBP feature is used to carry out rapid human head examination,and then using the head detector based on SVM and HOG features,the initial detection results are rediscriminated,which greatly reduces the error detection rate of the head target.In the multi-target tracking counting,this paper first studies the multi-target tracking algorithm based on path prediction.Because the crowd blocking problem is more serious in the oblique viewing angle,this paper uses the detected human head target instead of the traditional motion segmentation target to track.But the tracking effect of this method has a strong dependence on the detection of the human head target.At the same time,the detection of the human head target inevitably leads to missed detection and false detection,so the statistical accuracy is only 90%.Therefore,this paper studied the target tracking method based on kernel correlation filter,and proposed a multi-target tracking method combining kernel correlation filter and greedy algorithm.The experimental results show that the final tracking accuracy reaches 98.1%.Finally,based on the actual needs,this article design and accomplish a human traffic statistics software system.In crowd density estimation,this paper studied the latest progress in people counting and density estimation in the last two years and deeply studies on people counting and density estimation methods based on convolutional neural networks.According to the idea of differentiation,the method of making density map is simplified.Based on the idea of divide and conquer,a model of block convolutional neural network with multiple columns is proposed,which improves the training speed of the model and achieves good results.Finally,using the proposed method to discriminate the crowd density in different regions of the scene,the density distribution of the crowd in the monitoring visual field can be intuitively obtained.