Automatic segmentation of pancreas has always been a challenging problem in medical image segmentation. The pancreas is an organ with a high degree of anatomical variability, and it is difficult for the current multi-atlas segmentation methods to accurately segment the edges of the pancreas. Focusing on this problem, a segmentation algorithm is adopted based on multi-atlas registration to segment the pancreas, and optimizes a post-processing method of local dynamic threshold. In the label fusion stage, four label fusion algorithms are used for comparison: probability threshold fusion algorithm, Majority voting (MV) algorithm, STAPLE algorithm and SIMPLE algorithm. In the post-processing stage, the local dynamic threshold processing method is adopted. First, the target area is extracted from the target image through the preliminary segmentation result, and then the threshold value is automatically determined to realize the binarization of the area. Finally, the intersection with the preliminary segmentation result is taken as the final segmentation result. A leave-one-out cross-validation strategy was used to segment 80 NIH pancreatic CT images and 22 pancreatic CT images from local hospital at Shanghai, and the final DSC obtained were 79.98% and 81.30%, respectively. The experimental results show that the proposed method achieves effective segmentation of the pancreas.