Research And Implementation Of Dense Crowd Counting System Based On Deep Learning

With the accelerating process of urban construction,the continuous expansion of area and improvement of facilities,the rural population continues to flood to the city,which leads to places with dense crowds everywhere.Excessive density of crowds is prone to some accidents.The statistical analysis of the people under the surveillance video is carried out to realize the population statistics algorithm.It can provide effective technical means for optimizing the allocation of urban public resources,dispatching security personnel,preventing accidents and so on.In this paper,deep learning algorithm is adopted to detect and identify people in surveillance video,and statistical algorithm is added to carry out statistical calculation on the identified people,so as to realize the number of people under video surveillance.In this paper,a multi-scale fusion network algorithm based on deep learning is proposed to achieve the population statistics,and Open CV is used to obtain video data.The model algorithm in this paper is verified and analyzed on the existing three population count data sets.The multi-scale fusion network is used to replace the Fully connected layer network of the pre-trained VGG-16 network to learn and calculate the head probability value of each pixel in the density map,and then the total population estimated by the network is obtained according to the statistics of the probability value.The crowd density map shows the people in the image in the form of dots in the density map,so that the crowd distribution information can be more intuitively known.The multi-scale fusion network constructed in this paper is used to verify the effectiveness of the proposed method on three representative population image data sets.The experimental results show that the proposed method can effectively improve the accuracy of population statistics.In the Shanghaitech dataset,the accuracy of the network is improved compared with other population count networks,in which mean square error(MSE)of the dense population is reduced by 10%,and the mean absolute error(MAE)and mean square error(MSE)of the sparse population are both reduced by about 6%.Finally,a video data set is used to further verify the accuracy of the proposed method in the video crowd counting process,which indicates the effectiveness and robustness of the proposed method.The functions of the whole system in this paper are researched and developed under Windows 10 operating system.Taking Py Torch as the framework of crowd counting algorithm,using the C++ function library provided by Open CV to achieve video fetching and frame by frame output as images,using Open CV and the multi-scale fusion network algorithm of this paper as the main development technology to achieve various functions of the system background.Use the HBuilder X tool to complete the design and development of the front page.According to the different scenarios for different crowded degree requirements,this system provides a number of intensive threshold setting function,can set different values as warning value,system according to the result of counting and set threshold contrast,makes the corresponding prompt,and then the crowd statistical visualization display,convenient early warning information browsing and intensive number display.Finally,the function and system performance of each module of the system are tested by sampling the local video files to simulate the monitoring video.The test results show that all the functions of the system meet the requirements,and the recognition speed and accuracy are always maintained at a high level,showing a high stability,and at the same time has a certain expansible space.