Abstract: Crowd counting is challenged by many aspects, such as occlusion, perspective change and perspective effect, background imaging interference, etc. In this paper, we propose a multi-scale feature aggregation and dense connection network for crowd counting. A vital component of this network called multi-scale feature aggregation model (MFA) extracted multi-scale features through different convolution kernels and aggregated their cross-scale information for more accurate estimation. This model was tested on three public datasets Shanghai Tech A, Shanghai Tech B, UCF_QNRF and UCF_CC_50. The results show that the model has reduced the mean absolute error (MAE) and mean square error (MSE) to different degrees, and the model accuracy is better compared with CSRNet. In contrast with other models, the model makes full use of multi-scale feature information and improves the accuracy of the crowd counting.