Estimating the number of people in the surveillance scene is one of the important tasks of security monitoring. However it is difficult to estimate the number when the crowd is with clutter and severe occlusion. An improved crowd counting method based on the convolution neural network is proposed as for the number estimation under dense scenes. In order to reduce the effect of camera perspective distortion, the deep network and shallow network are used to extract the crowd characteristics, respectively. The convolution layers with different kernel sizes are also designed. Moreover, the extracted features are fused through a special structure with multi-scale extraction capability. The experimental results show that the crowd density map obtained by the improved network model is closer to the original scene information and the obtained prediction results are more precise.