【摘要】 本文的目的是尝试一种方法,可以准确地估计从任意图像与任意人群密度和任意角度人群计数。为此,我们应用简单而有效的多列卷积神经网络(MCNN)架构,将图像映射到它的人群密度图。允许在任意大小或分辨率模式下输入图像。通过利用不同大小感受野的过滤器,每列卷积神经网络学习特点是自适应人/头尺寸对应透视效果或图像分辨率变化。针对这具有挑战性的任务,我们进行了大量的实验,以验证所应用模型和方法的有效性。此外,实验表明,这个模型一旦训练另一个目标数据集,可以很容易地转移到一个模式相近的新应用领域。更多还原

【Abstract】 The purpose of this paper is to try a method that can accurately estimate population counts from arbitrary images with arbitrary population densities and arbitrary angles. To this end, we apply simple but effective multiple row convolutional neural network(MCNN) architecture to map the image to its population density map. Allows images to be entered in any size or resolution mode. By using filters of different size receptive fields, each column convolutional neural network learning feature is adaptive human/head size corresponding to perspective effect or image resolution change. In response to this challenging task, we conducted extensive experiments to validate the effectiveness of the model and method applied. Moreover, experiments show that once the model is trained, another target data set can easily be transferred to a new application field with similar patterns.更多还原