Crowd density estimation is a crucial component in surveillance systems to construct safe and efficient urban environments. Due to perspective distortion, individuals in crowd scenes diminish in size as they converge toward the vanishing point. Hence, there are significant visual variations in individuals’ size and appearance, which may lead to inaccurate estimations of crowd counts. This paper proposes an intuitive and effective loss function for the error estimation of crowd counts, particularly in vanishing crowd regions. Specifically, estimation errors in vanishing crowd regions are used to refine and generate network filters that are adaptive toward perspective distortion during network training. Extensive experiments on the challenging UCF-QNRF, WorldExpo, and ShanghaiTech benchmark datasets demonstrate the effectiveness of our novel loss function for training a network to achieve accurate crowd density estimation, particularly in the presence of perspective distortion.

人群密度估计是构建安全高效城市环境的监视系统中的关键组成部分。由于透视失真，人群场景中的个体向收敛点收敛时尺寸变小。因此，个体的大小和外观存在明显的视觉变化，这可能导致人群计数的估计不准确。本文提出了一种直观有效的损失函数，用于人群计数的误差估计，尤其是在消失的人群区域中。具体而言，消失的人群区域中的估计误差用于细化并生成网络过滤器，该网络过滤器可适应网络训练期间的透视失真。在具有挑战性的UCF-QNRF，WorldExpo，