Pedestrian detection is the key technology to realize intelligent traffic and passenger flow monitoring. Currently, the training model of deep learning method has achieved good results in pedestrian detection. However, when the training samples are poor, the training model often fails to achieve good results. In order to improve the effect of pedestrian detection under hazy weather and strong exposure environment, the dark channel defogging algorithm is applied to pretreat deep learning samples. And pedestrian detection model is trained with fast deep convolutional neural network. In this experiment, the dark channel defogging algorithm is applied to preprocess the 10,000 sample images. After that, the sample images preprocessed by the defogging algorithm with and without dark channel are used to train model, respectively. Finally, detection accuracy of these two models under different scenarios are compared. The experimental results show that the depth model obtained by using the dark channel defogging pretreatment sample has a better detection effect and the detection rate increases under many scenarios.