基于生成对抗网络动态建模的人群密度预测方法

Crowd density prediction method based on dynamic modeling of generative adversarial networks

为预测未来时刻人群密度图,对人群聚集提前预警,提出一种基于生成对抗网络动态建模的人群密度预测方法。采用生成对抗网络结构,生成器采用加入扩张卷积的U-Net网络捕捉人群空间分布信息,利用光流估计模型FlowNet提取人群运动信息,联合4项损失函数对人群空间和时序两方面约束。实验结果表明,Mall数据集上均方误差、峰值信噪比和结构相似性分别为30.97、24.26 dB和0.671,Airport数据集上分别为8.36、34.58 dB和0.931,较现有方法具有更好的性能,验证了该方法的有效性。

To predict the crowd density map in the future and give early warning of crowd gathering,a crowd density prediction method based on dynamic modeling of generative adversarial networks was proposed.The model was based on generative adversarial network structure,and the generator adopted a U-Net network incorporating dilated convolution to capture crowd spatial distribution information,and an optical flow estimation model FlowNet was adopted to extract crowd motion information,and a joint four-term loss function was used to constrain both spatial and temporal aspects of the crowd.Experimental results show that the mean-square error,the peak signal to noise ratio and the structural similarity index measure are 30.97,24.26 dB and 0.671 on the Mall dataset and 8.36,34.58 dB and 0.931 on the Airport dataset,respectively.It have better performance than the existing methods and the effectiveness of the method is verified.