Crowd density estimation is a critical task for crowd management and public safety. In this research, we propose a crowd density estimation method based on a cascaded multilevel convolution neural network. Our proposed method involves three steps: (1) Extracting characteristics from lower layers to high layers using a multilevel convolution neural network to enhance separability of crowd density characteristics. (2) Eliminating connections of redundant neurons in the convolution neural network based on similarity of a characteristic pattern in a down sampling layer, which speeds up characteristic extraction. (3) Training two multilevel convolution neural networks of different structures based on the difficulty level of the separability of crowd density samples, and connecting them in cascade based on sequences from simpleness to complexity to form a crowd density estimation model of the cascaded multilevel convolution neural network. Our proposed method achieves a better real-time effect compared with previous schemes in terms of detection accuracy.

Crowd density estimation, multilevel convolution neural network, separability, redundancy elimination, real-time detection.