基于卷积神经网络的多尺度融合特征图在人群密度估计中的应用

Application of Using Multiscale Fusion Feature Maps in Crowd Density Estimation Based on Convolutional Neural Network

提出了一种以Unet++为基础的卷积神经网络,适用于人群密度估计。该网络的优点是用并行连接的方式进行多尺度融合,结合浅层网络的细节信息和深层网络的高阶语义信息来消除两者之间过大的语义鸿沟。此外,还引入了膨胀卷积来提高网络性能。在Shanghai Tech和UCF_CC_50两个通用人群密度估计数据集上进行实验,选取平均绝对误差(MAE)和均方误差(MSE)作为评价指标。实验结果表明,在这两个数据集上该网络均有效降低了MAE和MSE,说明其在人群密度估计方面有较好的准确度和鲁棒性。

A convolutional neural network based on Unet++is proposed for crowd density estimation.The advantage of this network is to perform multi-scale fusion in a parallel connection,combining the details of the low-level network and the higher-level semantic information of the high-level network to eliminate the excessive semantic gap between the two,and also introduce dilated convolution to improve network performance.The experiment was conducted on two general population density estimation datasets,ShanghaiTech and UCF_CC_50,and the mean absolute error(MAE)and mean square error(MSE)are selected as evaluation indicators.Experimental results show that MAE and MSE are effectively reduced on these two data sets,indicating that the network has good accuracy and robustness in population density estimation.