Research Of 3D Point Cloud Instance Segmentation Algorithm Based On Topological Constraints

With the advancement of 3D perceptual devices,there is an increasing interest in perceptual tasks in the 3D domain.Point cloud-based research has become an important direction in the 3D field.In recent years,many excellent works on large-scale point clouds based on deep learning have emerged.However,the task of 3D point cloud instance segmentation is still challenging.The mining and constraining of instance-level geometric and topological information in 3D point cloud instance segmentation has been the focus of exploration in previous works.However,most of the existing algorithms are directly based on neural networks for information mining and supervised optimization.This will lead to the following problems:(1)The neural network approach is difficult to directly fit important topological features such as connected components,holes,and cavities of the instances,and the results of model segmentation tend to lose important topological structures.(2)These models lack instance-level supervision and perception in training and inference,making neighboring instances easily clustered together.To address the above problems,this paper proposes a 3D point cloud instance segmentation model based on topological loss,and proposes a simplified continuous homography module to optimize the topological loss,and the main innovative work and research results accomplished are as follows:(1)Propose a 3D point cloud instance segmentation model based on topological lossTo address the problem that neural networks are difficult to directly fit higher-order topological features,topological loss is introduced into the 3D point cloud instance segmentation model to improve the topological accuracy of segmentation results and enhance instance-level supervision and perception.In this paper,a two-stage instance segmentation network with point-level prediction and instance-level clustering is used to generate instances of point clouds,and a topological instance teacher network is designed to correct and guide the predicted results to improve the quality of the generated instances,and finally,a topological loss is proposed to supervise and constrain the important topological features of the predicted results to enhance the instance-level supervision of the network.(2)Proposed simplified continuous homodyne moduleTo address the problem of constrained efficiency of the model after the introduction of topological loss,a simplified continuous homotopy module is designed in this paper.This paper proposes a sampling method of scene pre-sampling mask for stable downsampling of instances in order to improve the stability of topological loss optimization.And based on the nearest neighbor perceptual field augmentation module,the optimization range of the loss is extended to improve the constraint effect of topological loss.Finally,a continuous cohomology network is designed to accelerate the efficiency of topological feature extraction by means of neural network prediction.The experimental results show that the model proposed in this paper has high accuracy and practicality in point cloud data instance segmentation tasks,and can be applied to various applications that require point cloud instance segmentation.