Abstract: Aiming at the requirement of single person pose estimation in multi-person scene, we combine YOLOv5 series model, object selection and lightweight light-duc model to estimate pose of specific person. YOLOv5 network was used to detect all people in the image. Meanwhile, the object selection composed of DeepSORT multi-object tracking algorithm and criteria-based selection was used to select specific person. The light-duc lightweight network was designed to estimate the pose of specific person. The experimental results show that compared with the original network, the speed of the proposed light-duc network is increased by 157%, combining YOLOv5s with light-duc model, the detection speed of single person video has great improvement, which lead to 319%.