In daily life, maintaining proper social distance can effectively stop the spread of epidemic viruses and help protect our lives and property. In this study, we developed a social distance detection system designed by a promoted yolov7 deep neural network, and ensured that the system distance error is fully applicable to daily pedestrian distance detection. Based on the faster and more accurate yolov7 framework, we first use a self-made specific dataset for data augmentation, then use k-means++ to design a pre-selected box that fits the pedestrian size, finally, use the SIOU loss function to replace CIOU to obtain faster convergence speed and optimization effect. The experimental results show that the detection speed of the network is 106 FPS and the accuracy value is 93.1%, which is 0.9% better than the original algorithm and achieves the effect of real-time pedestrian distance detection. Studying the imaging principle of monocular camera, we use Zhang’s calibration method to obtain the internal reference and aberration parameters of the camera, then use the algorithm of inverse perspective change to transform the region of interest into the viewpoint of the bird’s eye view, applying the scale factor in the viewpoint of the bird’s eye view can be converted into the specific distance of pedestrians in the actual three-dimensional space. The experiment was conducted in the entrance of the liberal arts building, and the calculated error was kept within 20mm, which ensured the availability of safe social distance.