MSIANet:多尺度交互注意力人群计数网络

MSIANet:Multi-scale Interactive Attention Crowd Counting Network

尺度变化、遮挡和复杂背景等因素使得拥挤场景下的人群数量估计成为一项具有挑战性的任务。为了应对人群图像中的尺度变化和现有多列网络中规模限制及特征相似性问题,该文提出一种多尺度交互注意力人群计数网络(Multi-Scale Interactive Attention crowd counting Network,MSIANet)。首先,设计了一个多尺度注意力模块,该模块使用4个具有不同感受野的分支提取不同尺度的特征,并将各分支提取的尺度特征进行交互,同时,使用注意力机制来限制多列网络的特征相似性问题。其次,在多尺度注意力模块的基础上设计了一个语义信息融合模块,该模块将主干网络的不同层次的语义信息进行交互,并将多尺度注意力模块分层堆叠,以充分利用多层语义信息。最后,基于多尺度注意力模块和语义信息融合模块构建了多尺度交互注意力人群计数网络,该网络充分利用多层次语义信息和多尺度信息生成高质量人群密度图。实验结果表明,与现有代表性的人群计数方法相比,该文提出的MSIANet可有效提升人群计数任务的准确性和鲁棒性。

Factors such as scale variation,occlusion and complex backgrounds make crowd number estimation in crowded scenes a challenging task.To cope with the scale variation in crowd images and the scope limitation and the feature similarity problem in existing multi-column networks,a Multi-Scale Interactive Attention crowd counting Network(MSIANet)is proposed in this paper.Firstly,a multi-scale attention module is designed,which uses four branches with different perceptual fields to extract features at different scales and interacts the scale features extracted from each branch.At the same time,an attention mechanism is used to limit the feature similarity problem of the multi-column network.Secondly,a semantic information fusion module is designed based on the multi-scale attention module,which interacts different levels of semantic information of the backbone network and stacks the multi-scale attention module in layers to make full use of the multi-layer semantic information.Finally,a multi-scale interactive attention crowd counting network is constructed based on the multi-scale attention module and the semantic information fusion module,which makes full use of multi-level semantic information and multi-scale information to generate high-quality crowd density maps.The experimental results show that compared with the existing representative crowd counting methods,the proposed MSIANet can effectively improve the accuracy and robustness of the crowd counting task.