Aiming at the problems of poor performance of traditional pedestrian detection and tracking algorithms and false detections in public places, crowded or complex background, this paper proposed an NMS algorithm based on the siamese network, which is to improve the detection and tracking of pedestrians in crowded accurately. This method uses less occluded visible parts to remove redundant detection boxes. In order to obtain the visible part, this paper proposed a double-box model (DBM) to predict the whole body part and the visible part of the pedestrian at the same time, which is to ensure the correspondence between the two boxes in the entire network, so as to achieve a better pedestrian detection task performance. This paper has carried out experimental verification on the CrowdHuman dataset, and the experimental results show that it has good robustness and detection accuracy for pedestrian detection in crowded and is better than other models by about 5%.