This study proposes a Poisson surface reconstruction algorithm based on the improved normal orientation to verify the inaccurate reconstruction results obtained because of the inconsistent normal orientation of the three-dimensional point cloud in the Poisson surface reconstruction process. First, the original point cloud is segmented into a number of cubic voxels using the bounding box and each valid voxel is marked as “position known” or “position unknown”. Then, a k-dimensional tree is used to reorder the point cloud in each valid voxel; the principal component analysis method is used for normal estimation. The new order is used as the normal orientation of propagation to realize the local orientation of the point cloud normal. For voxels marked as “position known”, the global orientation of the normal is completed according to the physical position of the voxel centroid. For voxels marked as “position unknown”, the global orientation of the normal is implemented according to the normal of the adjacent voxels. Finally, Poisson surface reconstruction is performed on the point cloud in combination with the normal orientation results. Experimental results demonstrate that the proposed algorithm can orient the normal of the complex surface point cloud to the same side and that the deviation of the normal estimation is approximately 0. Thus, the proposed algorithm can properly realize Poisson surface reconstruction and improve the accuracy of reconstructed surface.