针对密集人群场景存在的目标尺度多样、人群大尺度变化等问题,提出一种基于多尺度感知的密集人群计数网络.首先,考虑到小尺度目标在图像中占比较大,以VGG-16(visual geometry group 2016)网络为基础,引入空洞卷积模块,以挖掘图像细节信息;其次,为充分利用目标多尺度信息,构建新的上下文感知模块,以提取不同尺度之间的对比特征;最后,考虑到目标尺度连续变化的特点,设计多尺度特征聚合模块,提高密集尺度采样范围与多尺度信息交互,从而提升网络性能.实验结果显示:在ShangHai Tech(Part_A/Part_B)和UCF_CC_50数据集上,本文方法的平均绝对误差(mean absolute error,MAE)分别为 62.5、6.9、156.5,均方根误差(root mean square error,RMSE)分别为 95.7、11.0、223.3;相较于最优对比方法,在UCF_QNRF数据集上的MAE和RMSE分别降低1.1%和4.3%,在NWPU数据集上分别降低8.7%和13.9%.
Dense Crowd Counting Network Based on Multi-scale Perception
A dense crowd counting network based on multi-scale perception was proposed to solve the problems of diverse target scales and large-scale changes of crowds in dense crowd scenes.Firstly,since the small-scale targets account for a relatively large proportion of the images,a dilated convolution module was introduced based on the visual geometry group 2016(VGG-16)network to mine the detailed information in the images.Then,by utilizing the multi-scale information of the target,a novel context-aware module was designed to extract the contrast features between different scales.Finally,In view of the continuous change of target scales,the multi-scale feature aggregation module was designed to improve the sampling range of dense scales,enhance the interaction of multi-scale information,and thus improve the model performance.The experimental results show that mean absolute errors(MAEs)of the proposed method are 62.5,6.9,and 156.5,and the root mean square errors(RMSEs)are 95.7,11.0,and 223.3 on ShangHai Tech(Part_A/Part_B)and UCF_CC_50 datasets,respectively.Compared with the optimal method of comparison model,the MAE and RMSE are reduced by 1.1%and 4.3%on the UCF_QNRF dataset and by 8.7%and 13.9%on the NWPU dataset.
crowd density estimationmulti-scale aggregationdilated convolutiondensity map
国家科技期刊平台
NSTL
万方数据
