基于单列特征融合卷积神经网络的人群计数技术

Crowd Counting Technology Based on Single Column Feature Fusion Convolutional Neural Network

为了提高人群计数精度,设计一种基于单列特征融合卷积神经网络的人群计数技术。该网络包含前端网络、中端网络以及后端网络三个部分。前端网络采用VGG16网络的前十层并使用特征金字塔池化来融合基础特征;中端网络采用小尺寸卷积层堆叠以及特征融合的方法来提取多尺度特征;后端网络采用不同空洞率的空洞卷积层来提高感受野,从而得到人群密度图。实验在ShanghaiTech数据集上进行,结果表明,该网络在人群计数上具有良好的准确性与鲁棒性。

In order to improve the accuracy of crowd counting, a crowd counting technology based on single column feature fusion convolutional neural network is designed. The network includes three parts: front-end network, middle-end network and back-end network. The front-end network uses the first ten layers of the VGG16 network and uses feature pyramid pooling to fuse basic features;the mid-end network uses small-size convolution layers stacking and feature fusion methods to extract multi-scale features;the back-end network uses dilated convolutional layers with different dilated rates to improve the receptive field and obtain a crowd density map. The experiment was carried out on the ShanghaiTech dataset, and the experimental results show that the network has good accuracy and robustness in crowd counting.