Research On Instance Segmentation Algorithm Of Indoor Scene Based On 3D Point Cloud Data

Point cloud data is one of the most important types of 3D data.The point cloud data obtained by a three-dimensional scanner（such as a lidar scanner）is composed of many points like a cloud,and the attributes of each point generally consist of coordi-nates and color information.Compared with traditional two-dimensional pictures,three-dimensional point cloud data contains more information.The potential of point cloud data enables it to have a wide range of application scenarios,especially accurate point cloud segmentation,which has outstanding performance in 3D scene reconstruction,automatic driving,medical assistance,and other fields.Among them,the point cloud instance seg-mentation task of indoor scenes has significant research value.Indoor point cloud data can be obtained in real time at low cost through a portable device equipped with lidar.On the basis of using the point cloud instance segmentation algorithm to return the seg-mentation results in real time,functions such as specific object search and path guidance can be realized.The point cloud instance segmentation algorithm for indoor scenes has significantly improved based on semantic prediction,point cloud voxelization,and other technologies.However,there are still key issues that need to be solved urgently,such as point cloud instance segmentation algorithms generally rely too much on semantic pre-diction,and point cloud acquisition equipment cannot obtain complete scene data.This thesis analyzes the impact of the above problems on point cloud instance segmentation in detail,and proposes targeted solutions.The main content of the thesis is as follows:（1）A point cloud instance segmentation model based on semantic error correction is proposed.More accurate semantic features are obtained by hard sample mining and re-extracting their semantic features based on adding more accurate local semantic features.At the same time,the model abandons the general clustering method that takes semantic prediction as a mandatory premise and designs a novel soft clustering mechanism,which calculates the similarity according to the distance and semantics between points and uses the similarity to complete the final point cloud aggregation.The error correction ability for semantic errors is realized,and the model achieved 70.0%on the authoritative indoor point cloud dataset in terms of₅₀.（2）A point cloud instance segmentation model based on point cloud completion is proposed.The incompleteness of point clouds in indoor scenes is inevitable.This model combines point cloud completion to predict the overall point cloud based on part of the point cloud.A new completion and merging module are designed.Perform point cloud completion on the initially obtained point cloud instance sets,then calculate the similar-ity between the completed instance sets,and finally merge the instance sets with higher similarity.From the perspective of point cloud completion,the problems of point cloud in-completeness and excessive instance segmentation are optimized,and the model achieved70.7%on the authoritative indoor point cloud dataset in terms of₅₀.（3）An indoor target positioning and navigation applications based on point cloud instance segmentation is designed and implemented.Object seeking in indoor scenes has great potential applications for the visually impaired.With the point cloud instance seg-mentation model proposed in this thesis and related algorithms such as speech recogni-tion and pathfinding algorithms,this application can help the visually impaired people use voice interaction to achieve the positioning and navigation of target objects in indoor scenes,which proves the practical value from the perspective of application.