基于多尺度全卷积网络特征融合的人群计数

Crowd Counting Based on Feature Fusion of Multi-Scale Fully Convolutional Networks

图像中的人群计数在公共安全领域具有重要价值.为了解决由于摄像机透视效果、人群密度分布不均匀和严重遮挡等导致人群计数准确率低的问题,本文提出一种多尺度全卷积网络架构,用于准确地估计任意摄像头视角的静态图片的人群密度.通过利用不同尺度的卷积核,使分支网络能更好地学习图像中头部特征变化.同时,由于每个分支网络设计的网络层数量不同,因此这种多尺度的网络组合能够有效地捕捉高层的语义信息和低层的细节信息.实验结果显示,本方法在Shanghai-tech标准数据集上具有较高的人群计数准确率.

The crowd counting in the image has an important value in public safety field.In order to solve the problem of low accuracy of population counting due to camera perspective effects,uneven population density distribution,and severe occlusion,this paper presents a multi-scale full convolutional network architecture to accurately estimate the crowd density of static images from arbitrary camera perspectives.Different scales of convolution kernels helps the branch network to learn the head feature changes in the image well.At the same time,there are differences in the number of network layers per branch network,so the combination of multi-level networks can effectively capture high-level semantic information and low-level detailed information.The results of experiments prove that this method has achieved high accuracy of crowd counting based on the Shanghai-tech standard data set.