Abstract: In response to the deficiencies such as big error and poor performance in the traditional
method of real-time crowd density estimation, a new one based on CNN is proposed. By comparing
the accuracy and real-time of four common network structures—AlexNet, VGGNet, GoogLeNet, and
ResNet, the GoogLeNet which has relatively better comprehensive performance is chosen as the
model for crowd density estimation. We used the key-frame extraction technology to realize real-time
crowd density estimation and briefly analyze the crowd density feature map. Finally, examples are
analyzed to verify the real time, accuracy, and feasibility of this new method of real-time crowd
density estimation.

Key words: crowd density, convolutional neural networks, video processing, real-time estimation