Deep learning methods have proven to be effective in the field of crowd analysis recently. Nonetheless, the performance of deep learning models is affected by the inadequacy of training datasets. Because of policy implications and privacy restrictions, crowd data is commonly difficult to access. In order to overcome the difficulty of insufficient dataset, the previous work used to synthesize labelled crowd data in outdoor scenes and virtual games. However, these methods perform data synthesis with limited environmental information and inflexible crowd rules, usually in unauthentic environment. In this paper, a tool for synthesizing crowd data in BIM models with multiple scenes is proposed. This tool can make full use of the comprehensive information of real-world buildings, and conduct crowd simulations by setting behavior rules. The synthesized dataset is used for data augmentation for crowd analysis problems and the experimental results clearly confirm the effectiveness of the tool.