一种面向人群计数的卷积注意力网络模型

Convolutional Attention Network for Crowd Counting

随着当今国际社会形势逐渐复杂,公共安全和社会稳定面临严峻挑战。视频监控作为维护社会安定与建设智慧城市的重要手段,广泛应用于城市安全管理。高效的人群计数是实现基于视频进行安全管理的一个难点问题,旨在分析计算视频或图片场景中的人数。人群计数对控制关键场所人数、指挥公共交通、控制疫情蔓延、保障社会稳定具有重要积极意义。然而,人群计数问题仍然存在背景干扰、目标遮挡、目标尺度不一和目标分布不均等挑战,导致计数准确度较低。为了解决这些问题,梳理了人群计数发展的时间线,分析了现有方法的不足,并针对这些不足提出了基于相似性度量的卷积注意力网络。该方法结合基于相似性度量的损失函数和基于注意力机制的卷积神经网络模块,有效缓解了人群计数中背景干扰、目标遮挡、目标尺度不一和目标分布不均四个问题。通过在数据集上的实验和相关对比分析发现,基于相似性度量的卷积注意力网络具有很好的准确性和稳定性。

With the increasingly complex international situation,public security and social stability are facing severe challenges.As an important means to maintain social stability and develop smart cities,video surveillance has been widely used for managing public security.Crowd counting is a critical issue in video surveillance-based security management and it aims to calculate the number of people in the video or picture scene.It is of high significance for controlling the number of people in key places,directing public transportation,controlling the spread of the epidemic,and ensuring social stability.However,there are still some challenges in crowd counting,such as background interference,target occlusion,target scale difference and target distribution inequality.In order to solve these challenges,this paper reviews the timeline of crowd counting development,analyzes the shortcomings of existing methods,and puts forward corresponding solutions.Then,the paper proposes a convolutional attention network based on similarity measure,which combines the loss function based on similarity measure and the convolutional neural network module based on attention mechanism,and effectively alleviates the four issues of background interference,target occlusion,target scale difference and uneven distribution in crowd counting.Through the experiments and comparative analysis on datasets,it is found that the convolutional attention network based on similarity measure achieves good accuracy and stability in crowd counting.