一种多列特征图融合的深度人群计数算法

A Deep Crowd Counting Algorithm Based on Multi-column Feature Map Fusion

针对复杂开放环境下人群密度估计中的多尺度目标和小目标感知问题,提出了一种基于特征图融合的多列卷积神经网络的人群密度估计算法.所提出的特征图融合方式,一方面综合利用高层语义特征与底层细节特征,实现了对小目标的感知;另一方面大幅提高基础网络集成数量以应对目标多尺度问题,最终提高了人群密度估计的准确性.实验结果表明,所提算法有效提高了密集开放场景中人群计数的准确性.

Aiming to solve the multi-scale targets and small person recognition problem of crowd counting in open scene,a feature map fusion convolutional neural network to promote the accuracy of density map regression was proposed.The feature map fusion had the two advantages.Firstly,it could combine the low-level detailed features with high-level semantic features,which promoted the detection of small person.Secondly,networks were assembled in the overall network by combining the feature maps,which largely promoted the network′s performance on recognizing multi-scale targets.The experimental results demonstrated the accuracy promotion of the proposed method in crowded open scenes.