基于上下文特征重聚合网络的人群计数

Context-aware feature reaggregation network for crowd counting

针对计数问题中人群目标尺度的变化问题,提出了一种基于上下文特征重聚合的计数算法。将高层网络提取的语义信息与底层网络提取的人群尺度细节信息相结合,旨在利用浅层网络中提取的信息向深层网络提取的特征中融入不同尺度的行人目标特征,从而融合多种尺度的人群特征回归出高质量的人群密度图。此外,在ShanghaiTech、UCF_CC_50以及UCF_QNRF三个数据集进行算法的性能验证,并通过结构实验验证本文结构的有效性。

Aiming at the problem of the target scale-aware in crowd counting,this paper proposes a counting algorithm based on the reaggregation of contextual features.The semantic information extracted by the high-level network is combined with the crowd-scale detail information extracted by the low-level network,and the purpose is to use the information extracted from the low network to integrate the features of different scales into the features extracted from the deep network,thereby fusing multiple scales crowd feature return to a high-quality crowd density map.In addition,this paper performs algorithm performance verification on ShanghaiTech,UCF_CC_50 and UCF_QNRF datasets and the effectiveness of structures is verified through structural experiments.