Abstract: Against the problem of large public venues evacuation risk assessment, a deep learning evaluation model for emergency evacuation that integrates generative adversarial networks (GAN) and convolutional neural networks (CNN) was proposed. The problem of insufficient evacuation data is solved by data enhancement through WGAN (Wasserstein GAN). Based on CNN, two network structures of LeNet and ResNet were used for data training. A large gymnasium was taken as an example, to perform the evacuation risk evaluation by the method. The research results showed that an effective risk assessment model can be established to achieve rapid risk assessment for emergency evacuation.

Key words: emergency evacuation, generative adversarial networks, convolutional neural networks, risk assessment