Research On Dense Population Statistics Algorithm Based On Convolutional Neural Network

In the wake of the rapid development of social economy,high-tech electronic products emerge in multitude.Video surveillance devices can be seen everywhere in public areas.Especially in the priority control regions with dense population and large flow,it can replace the personnel on duty in a gradual way and realize 24-hour reliable monitoring.For ensuring municipal security protection,it provides effective technical support for safeguarding urban safety and accelerating the construction of smart cities.With the development of deep convolutional neural networks,the existing dense population statistics algorithms have achieved significant results,which can automatically analyze and calculate the crowd density in the scene,and realize the statistics and positioning of the number of people in the major areas of a specific scene.However,there are still some insurmountable challenges:(1)By means of comparison and classification of the existing algorithms,the conclusion can be obtained that the detection based method can accurately locate the target.However,the disadvantage is that the high occlusion problem is severe,which is only applicable to the sparse crowd;The method based on density regression has an impressively high accuracy,but its disadvantage shows as its inability to locate accurately,meanwhile its sensitivity to the resolution of the input image is extremely high.(2)Deep Convolutional Neural Network requires,in general,supplementary computation,extra memory as well as power consumption.With the deepening of network depth,there is also a lot of computational redundancy,making it difficult to deploy on edge devices and unable to guarantee real-time performance.In summary,to address the shortcomings of existing algorithms,this paper proposes the corresponding convolutional neural network-based statistical calculation method for the number of people,using a target detection-based method and a density estimation-based method,respectively,with improvements in terms of both accuracy and speed.The main research work and contributions of this paper are summarized in two parts:1.Target detectionIn order to improve the accuracy and speed of people counting in video monitoring and solve the problem of high occlusion caused by clothing blocking in traditional human body detection,an improved lightweight head detection MKYOLOv3-tiny is proposed.This method optimizes the network structure of YOLOv3-tiny,and performs multi-scale fusion of low-level human head features to achieve classification prediction and position regression of different convolutional layers,thereby improving the accuracy of detection;Considering the characteristics of small head and the concept of an effective receptive field,and K-means clustering reduces the size of the initial candidate frame and improves the accuracy of the candidate frame.The improved model,compared with the original method on the Brainwash dense head detection dataset approved by experiment,improves the accuracy and reduces the rate of missed detection.2.Density estimationTo solve the problem of dense and uneven flow distribution and complex computation,a crowd counting network framework based on deep separable convolution is designed.The framework consists of a front-end network along with a back-end network.The front-end network extracted the low-level features through VGG19,and introduced deep separable convolution to reduce the computation and enhance the feature extraction ability of the model.Back-end network using convolution of different expanding rate characteristics of normalized processing,to ensure the continuity of the information at the same time increase the receptive field.Finally,a comparative experiment is conducted on several current mainstream population counting data sets,and the experimental results show that the proposed algorithm is effective in population counting.