The existing crowd counting methods are not suitable for the subway scene. Therefore, a crowd counting model based on convolutional neural network is proposed. The model takes the VGG16 as the front-end network to extract the shallow features, and an M-Inception structure is combined with the dilated convolution to form the back-end network, which can increase the receptive field and adapt to different sizes of pedestrian targets at different monitoring angles. And a weighted loss function combining the head count loss and density map loss is proposed. The proposed algorithm is compared with four existing models. The experimental results show that the Mean Absolute Error(MAE) and Mean Square Error(MSE) of the proposed algorithm are 1.46 and 2.13, better than those of the four comparison models. The proposed model is deployed to Hisilicon embedded chip. The test results show that the proposed model can achieve high computing speed and accuracy on the embedded chip, which can meet the requirements of the actual application scenarios.