Research And Implementation Of Crowd Counting Algorithm Based On Video Surveillance

Crowd counting is an important part of intelligent surveillance systems,which play an increasingly important role in areas such as public safety,smart transportation and smart business.With the continuous development of convolutional neural networks in crowd counting tasks,the accuracy of counting has been continuously improved.However,in real scenarios,the existing algorithms still face many challenges.In this paper,the proposed algorithm and the implementation of the system will be studied,and the main work is summarized as follows.Crowd counting estimates the number of people in video surveillance images.In this paper,a crowd counting network with cross-layer connections as the backbone of the network is proposed to solve the problem of population scale changes.Compared with the method of increasing the number of different columns of the network,the cross-layer connected network structure can increase the number of perceptual fields of different sizes when the network depth is the same.It realizes counting by integrating and summing the crowd density map obtained from the network regression.However,the actual population is unevenly distributed,which affects the accuracy of counting.To overcome this problem,this paper introduces a spatial attention network at the back end of the network,which can dynamically adjust the density estimate of each corresponding region by learning the scaling factor to generate a more accurate local density map,thus reducing local errors.Extensive experiments were conducted on the Shanghai Tech,UCF-QNRF,UCF＿CC＿50,and NWPU-Crowd datasets.Compared with other methods,the proposed method is able to improve the counting accuracy when the population scale changes,and reduce the counting error when the population is unevenly distributed.In order to implement the system,this paper first analyzes and designs the device architecture and functional modules from the system requirement analysis.On the basis of building the hardware devices and development environment related to system development,the assignment and implementation of software functional modules are then carried out.Finally,the overall implementation of the system was carried out by combining the hardware devices and software module calls.Two different indoor scenarios,a university building and a subway station interchange,were selected for the actual system verification.The actual test results proved that the processing speed for standard surveillance video images can reach 12frames/second and the statistical accuracy can reach more than 85% under the hardware environment of this system.The implementation of the crowd counting system proves the effectiveness of the method,and the acquisition of the number of people information can also provide data support for the subsequent crowd detection.