With the economic growth, the number of rail transit passengers increase dramatically. And the related safety problems cause many laboratories study the use of informationalized ways to identify passenger flow. The existing detection methods have the problems of mis- and false-detection to varying degrees under the situations of dense passenger flow, dim light and mutual blockage between pedestrians. In recent years, deep learning technology has become increasingly capable of image detection. Thus, this paper proposes a strategy of applying deep learning to detect the subway passenger flow. However, deep learning techniques require a great deal of data sample to train the deep convolution network, and the exist database are mostly single targets which cannot meet the requirements of dense passenger flow detection. In order to solve this problem, this paper collects the specific subway passenger video for monitoring and use the manual annotation method to make a large number of dense data sample for deep convolutional network. We calibrated this data sample twice with different calibration methods. The second calibration filtered the excess calibration target background as well as improved the missing calibration box. Then we chose the Faster R-CNN target detection framework in combination with the VGG-16 and ZF convolution neural networks to experiment on these two data sample. Finally, we compare the results of different models with the same calibration methods and different calibration methods with the same model. And the best model shows good generalization ability.