| Scene flow is a 3D motion field defined between consecutive scenes.Fast and accurate estimation of scene flow can obtain motion information between consecutive scenes,which is of great importance for robotics and autonomous driving tasks.Traditional methods for point cloud scene flow estimation have been suffering from time-consuming and insufficient accuracy,while deep learning-based methods for point cloud scene flow estimation have shown much better performance.In this area,there are still many challenges in regression of scene flow from consecutive point clouds.We simply divide the scene flow task into two processes: matching and regression.The problems in the matching process include increasing the matching range and improving the matching accuracy,etc.;the problems in the regression process include retaining detailed information in the upsampling process,outlier processing,smoothing optimization,etc.Many current point clouds scene flow estimation methods focus more on the matching process and less on the regression process,especially on the optimization of scene flow based on geometric knowledge.In this paper,we propose a point clouds scene flow propagation update method based on neighborhood consistency,which uses a priori point clouds geometric information in the regression process for scene flow optimization.Specifically,the neighborhood consistency based point clouds scene flow update method includes a confidence prediction module and a scene flow propagation module: the confidence prediction module estimates a confidence map based on neighborhood information for the coarse scene flow graph estimated by the backbone network,which indicates the accuracy of the scene flow estimation at each point on the point cloud.The scene flow propagation module updates the scene flows of the low confidence point set according to the geometric constraint of local consistency,thus improving the accuracy of the whole scene flow graph.The test evaluation results on synthetic and real datasets demonstrate the effectiveness of the method and the improvement is more obvious on real datasets because the geometric assumption of neighborhood consistency is more consistent with the priori assumption of real scenes.Since labeling scene flow vectors for each point in a point cloud scene is very expensive,real point cloud scene datasets have almost no scene flow labels,and existing point cloud scene flow supervise methods are mainly trained on synthetic datasets and fine-tuned and tested on real datasets.This limitation of training data makes the training of the network without the participation of real scenes,which limits the scene flow estimation network in real environments effectiveness of the network.To solve this problem,this paper proposes a semisupervise scene flow estimation method,which uses the output results of two scene stream networks with the same structure but different initialization to construct pseudo-labels,and uses the pseudo-labels to supervise the unlabeled data.The semi-supervise training approach can effectively use a small amount of existing labeled real point cloud scenes while training the network with a large amount of unlabeled point cloud data.The test results on real datasets significantly outperform the results of fine-tuning using only labeled data,demonstrating great promise for application in real scenes. |
