Abstract: Scene flow is a 3D motion field between continuous dynamic scenes, which is widely applied in robotics and autonomous driving tasks. Existing methods ignore the correlation of point cloud points and focus only on the point-by-point matching relationship between the source point cloud and target point cloud, which is still challenging to estimate the scene flow accurately at the points with insufficient local feature information, because the matching relationship depends entirely on the feature information of the point cloud data. Considering the correlation property of the source point cloud’s local regions, the NCPUM (neighborhood consistency propagation update method) is proposed to propagate the scene flow from high-confidence points to low-confidence points in local regions, so as to optimize the scene flow at the points with insufficient local feature information. Specifically, NCPUM consists of two modules: the confidence prediction module, which predicts the confidence of the source point cloud according to the priori distribution map of scene flow; the scene flow propagation module, which updates the scene flow of the low confidence point set based on the local area consistency constraint. We evaluate NCPUM on both challenging synthetic data from Flyingthing3D and real Lidar scans from KITTI, and experiment results outperform by previous methods a large margin in accuracy, especially on KITTI dataset, because the neighborhood consistency is more applicable with the a priori assumptions of real Lidar scans.