Small pedestrian targets at long distance have problems of few pixels and lack of texture information. The deep convolutional neural network is difficult to extract fine-grained features of small objects. This article proposes a long-distance pedestrian small target detection method. Firstly, based on the YOLOv4 algorithm, the shallower features are introduced to improve the feature pyramid to extract fine-grained features of pedestrian small objects. An adaptive feature fusion layer is proposed based on the gravity model to increase dependency between multiple feature layers and prevent the loss of small target feature information. Then, ESRGAN is utilized to increase pedestrian small target features number and improve pedestrian small target detection accuracy. Finally, the small pedestrian targets are selected with a proportion of 0. 004% ~ 0. 026% in the image pixels to establish the self-made data set. Compared with Faster RCNN, ION, and YOLOv4, the mAP0. 5 values are increased by 25. 2% , 26. 3% and 11. 9% . And the FPS reaches 24. The research results have important application value in the field of long-distance security monitoring