Large scale crowd analysis based on convolutional neural network

• A method to estimate the number of crowd flow with CNN models is proposed.
• A database with 140 thousand samples from real scenes is build.
• The experiments perform robust under various scenes, weather or crowded condition.

Nowadays crowd surveillance is an active area of research. Crowd surveillance is always affected by various conditions, such as different scenes, weather, or density of crowd, which restricts the real application. This paper proposes a convolutional neural network (CNN) based method to monitor the number of crowd flow, such as the number of entering or leaving people in high density crowd. It uses an indirect strategy of combining classification CNN with regression CNN, which is more robust than the direct way. A large enough database is built with lots of real videos of public gates, and plenty of experiments show that the proposed method performs well under various weather conditions no matter either in daytime or at night.