基于深度特征融合生成的密集人群计数网络

为了进一步提高密集人群计数任务的计数精度,提出一种利用深度语义特征逐步降维重建的密集人群计数网络。前端采用深度卷积网络得到基本的深度语义特征;后端采用基于空洞卷积的多尺度特征融合块来丰富深度语义特征。通过语义重建块与上采样相结合,在进行多次降维重建以后生成与原始图像相同分辨率的人群密度图,并由此得到人群数量。将该模型在公开的数据集ShanghaiTech、UCF_CC_50、UCF-QNRF上与历年的主要方法进行对比,该方法无论是在人群计数精度还是密度图质量上都体现出了明显的优势,同时在多个数据上的验证实验表明模型具有较好的鲁棒性。

DENSE CROWD COUNTING NETWORK BASED ON DEPTH FEATURE FUSION

In order to further improve the counting accuracy of the intensive population counting task,this paper proposes a dense population counting network that uses the deep semantic feature to gradually reduce the dimensionality reconstruction.It was mainly composed of two parts.The front end adopted the deep convolutional network to obtain the basic deep semantic features.At the back end,the multi-scale feature fusion block based on the dilation convolution was used to enrich the deep semantic features.Through the combination of the semantic reconstruction block and the upsampling,the population density map with the same resolution as the original image was generated after performing multiple dimensionality reconstruction,and the number of people was obtained.The proposed model was compared with the main methods of the past years in the open datasets ShanghaiTech,UCF_CC_50 and UCF-QNRF.It shows obvious advantages both in crowd counting accuracy and density map quality.Validation experiments on the data show that the model has good robustness.