论文《Dense crowd counting from still images with convolutional neural networks》

创新点:使用深度学习方法估计一张图片中中等、高度拥塞程度的人群

解决问题

    人群计数算法的一些局限性

    (1)当人群规模达到成百上千时,这些算法只能用来估计人群密度而不能计算人群总数;(2)通常依赖于人群视频中的时序,但是这点在静止的图片中是不适用的。

人群没有定义好的形状,这点让人群特征提取变得困难;怎么提取能够表示人群信息中的特征也是很重要的;

提出方法

这篇文章提出了一个简单和一般的区别的基于学习的框架;提供AHU-CROWD数据集覆盖不同的场景,所有的人在数据集中都可以区分出来;

目的

使用在卷积网络中学习的框架向量在每个局部区域估计人群总数;并希望我们估计得数目和GT之间得偏差尽可能小;

Related Work:

人群计数方法分为两类:检测计数回归计数

回归计数:不通过分割或检测个体来学习低级特征和人计数之间的直接映射关系;

Methodology:

论文理解《Dense crowd counting from still images with convolutional neural networks》

论文理解《Dense crowd counting from still images with convolutional neural networks》

论文理解《Dense crowd counting from still images with convolutional neural networks》

上式是对第K块提取的特征向量;

论文理解《Dense crowd counting from still images with convolutional neural networks》

上式是对每个图像块估计的人数;

论文理解《Dense crowd counting from still images with convolutional neural networks》

上式是人群密度分类损失;

论文理解《Dense crowd counting from still images with convolutional neural networks》

下图是人群计数学习算法:

论文理解《Dense crowd counting from still images with convolutional neural networks》

EXPERIMENTS:

提出的网络架构:

论文理解《Dense crowd counting from still images with convolutional neural networks》达到了最好的分类准确率;

使用两个评判标准:

论文理解《Dense crowd counting from still images with convolutional neural networks》绝对误差和相对误差;

实验结果:

we use Single Signal Crowd ConvNet Model (SSCCM) trained on AHU-CROWD and finetune it with the UCSD training data.

论文理解《Dense crowd counting from still images with convolutional neural networks》

CONCLUSION

prove that our proposed architecture outperforms both traditional methods based on head detection and learning based methods on UCSD [15], and achieves the state-of-the-art result on UCF-CROWD .

Potential Direction

To combine temporal information and multiresolution information with ConvNet can be explored in our future work, to make a better estimation and apply this deep learning architecture in video surveillance and safety management




相关文章: