The Crowd Counting Model Based On Attention Mechanism

Large-scale crowd gatherings in public places are likely to cause huge potential safety hazards.If they cannot be evacuated in time,trampling accidents are likely to occur.Crowd counting and density estimation have always been research hotspots in the field of visual surveillance,which can help analyze 'the dynamic changes in the distribution of people in the scene and provide early warning of potential risks.Therefore,related algorithms are widely used in the monitoring systems of public places such as airports,stations,subways,shopping malls,etc.,and are of great significance to the field of public safety.Crowd counting and density estimation are extremely challenging topics.Factors such as light changes,background interference,viewing angles,and occlusion all increase the difficulty of handling the problem.The traditional counting method is based on pedestrian detection and is usually suitable for scenes where the crowd density is relatively sparse.In recent years,with the continuous development of deep learning and computer vision technology,the analysis of dense and complex scenes has received more and more attention.This paper proposes a complex crowd counting model based on attention mechanism.The model is mainly composed of attention network and crowd density network.The attention network is used to locate the potential position of the crowd,distinguish the foreground area from the background area,and overcome the interference caused by the background area to the counting.The crowd density network is used to generate a preliminary population density map.The attention map and the group density map are masked to obtain the refined population distribution.Finally,through the image integration operation,the crowd is counted.In addition,we also introduced the extended convolution operation to a certain extent to solve the problem of multi-scale crowd detection.We verify the effectiveness of the proposed model on two standard datasets,ShanghaiTech and UCF_CC₅0.These two datasets contain scenes with dense crowds in real environments.The experimental results show that:?1?the network fusion strategy based on extended convolution can effectively solve the problem of crowd multi-scale detection;?2?the attention network can reduce the background interference to the crowd counting accuracy and generate a high-quality crowd density map.Compared with standard baseline methods,the proposed model has improved accuracy in population density estimation and population counting.