Research On Crowd Counting Algorithm Based On Multi-scale Perception Attention Network

With the rapid development of social economy,people's collective activities are also increasing,and crowd trampling incidents sometimes occur due to crowd congestion.Therefore,it is an important task in the field of intelligent video surveillance system,to make accurate estimation of people in public places to prevent the occurrence of dangerous events.With the continuous development of computer vision technology,the accuracy of crowd counting has been greatly improved,and there are still many challenges in practical application.The so-called crowd counting is order to analyze and calculate the distribution characteristics of the crowd,through deep learning the mapping relationship between the image and the image density map,and then achieve the effect of crowd density estimation.There are mainly two challenges in crowd counting research.One is that the multi-scale features in the actual image set cannot be accurately extracted,which leads to the low accuracy of counting.The second is the misjudgment result of the crowd target due to the complex background in the crowd image.In view of the above two problems,this paper mainly studied the crowd counting algorithm,the research contents and innovation points are as follows:1.In view of the problem that the multi-scale features in the image,which cannot be fully extracted,the traditional convolutional neural network will extract the global features of the image,and then conduct the poolling operation of sampling on the input image set,which will lose the information of part of the population.In order to solve this problem,this paper adds the transposed convolution operation in the multi-column network structure,to ensure the same size of the merged features and make up for the loss of information details caused by previous pooling.2.A convolutional neural network structure,which is composed of convolutional branches and channel attention units,is designed to expand the scope of the receptive field.At the same time allow large,medium and small three sizes of input image resolution,which can effectively deal with the image in the man's head size differences,further enhance network perception of people under different scales.3.The attention mechanism is introduced into the network.In order to share it with the parameters in the crowd counting task when paying attention to the position of the head in the space,effectively filtering the information of the non-head in the image,and solving the problem of misjudgment of the target in the complex background of the image crowd.In the process of counting,the method of density map is adopted to avoid complex operations such as foreground segmentation and head extraction during image processing.In order to verify the effectiveness of the algorithm,a series of experiments were performed based on three international public data sets such as ShanghaiTech,UCF_CC₅0 and WorldExpo '10.The experimental results show that the proposed method is more suitable for the multi-scale crowd scenes and crowd counting under complicated background.The algorithm model also has good adaptability in this paper.The proposed algorithm is confirmed by comparing the current mainstream algorithms.It is better than some current mainstream algorithms and it has also greatly improved the accuracy of counting.The designed network has strong robustness.