基于背景辅助的高效人群计数多任务学习网络

An Efficient Background Assistance Based on Multi-Task Learning Network for Crowd Counting

在人群计数领域中,复杂背景干扰一直是一个具有挑战性的问题.现有研究通过引入注意力机制等方式弱化背景噪声对计数的影响.但是,随着研究的深入,人群计数网络规模不断扩大,影响了计算效率和实时应用.为了解决复杂背景问题并提高计数效率,该文提出了一个基于背景辅助的高效人群计数多任务学习网络(BAMTLNet).与现有网络不同,为了减少网络的参数量,只采用了VGG-16的前7层作为前端网络.在后端网络中,为了解决复杂背景问题,我们使用了两个高度相关的人群任务:①生成估计密度图主任务,采用3个普通卷积层生成密度图,通过积分获得单张图片的人数.②复杂背景分割辅助任务,采用3个特定的膨胀卷积层生成图片的背景分割图.两个任务直接连接在前端网络后,没有相互交叉.我们还设计了背景辅助多任务损失函数,通过硬参数共享的方式优化前端网络参数,向主任务传递复杂背景的高级语义信息并优化网络.该端到端人群计数多任务学习网络仅有10层卷积层,参数量小,实现了网络轻量化.在3个人群计数基准数据集上进行了实验,获得了令人满意的结果.

Complex background interference is still a challenging issue in crowd counting.In the existing crowd counting methods,attention has been paid on other approaches utilized to reduce the influence of background.As the research continues,the scale of crowd counting networks is growing,which makes a negative influence on computing efficiency and real-time application.Therefore,to solve complex background problem and to improve the counting efficiency,an efficient background assistance based on multi-task learning network(BAMTLNet)has been proposed in this paper.Unlike the existing networks,the first seven layers of VGG-16 has only been used as the front-end network to reduce the number of network parameters.For the problem of complex background,two highly correlated crowd counting tasks have been utilized in the back-end network:1)The main task of generating estimated density map,which adopts three general convolutional layers to generate a density map,and obtains the number of people in a single image by integration.2)The auxiliary task of complex background segmentation,by which to use three specific dilated convolutional layers to generate a background segmentation map.The two tasks have directly been connected behind the front-end network with no crossing.Besides,a background-assisted multi-task loss function has been designed to optimize the front-end network parameters through hard parameters sharing,by which to transfer the high-level semantic information of complex background to the main tasks and to optimize the network.This end-to-end crowd counting multi-task learning network is able to achieve comparable performance with only ten convolutional layers and less parameters.extensive experiments have been conducted on three crowd counting benchmark datasets and obtain satisfactory results.