RGB-D Scene Segmentation Method Based On Multi-feature LightGBM

Scene segmentation of a point cloud refers to the task of identifying a portion of an object corresponding to a single semantic unit in point cloud scene data.Point cloud data with similar properties is mainly clustered into connected regions by features such as color,depth value and normal vector to achieve separation of individual objects in the scene.It is a key step and important research topic in the field of machine vision applications such as target detection,3D reconstruction,scene understanding,and object recognition.In the real world,due to the complex and diverse backgrounds,stack occlusion phenomena between objects,and sensor noise interference,the research work on segmentation problems in point cloud scenes has always been very challenging.Because of the existing indoor small scene RGB-D point cloud data scene segmentation technology has low segmentation accuracy and is susceptible to sensor noise.In this paper,we designed A scene segmentation algorithm based on multi-feature iteration LightGBM classifier and morphology optimization algorithm.The algorithm achieves the segmentation of point cloud data in the desktop scene,and lays a solid foundation for the research of indoor scene recognition and registration.Firstly,we use Matlab to extract the features of the original RGB-D point cloud data,calculate the color,depth,normal vector and curvature of the RGB-D point cloud.Among them,the color features and the normal vector features are three dimensions,depth and Curvature features are one dimension.The original RGB-D point cloud data is processed into eight-dimensional feature data,and then the domain features of each point are combined by a sliding window algorithm.Because of the sampling according to the phenomenon that the sample ratio is too large compared with the non-boundary points.The method is to constructan efficient training subset and test subset for the prediction of the edge contour of the next object.Secondly,according to the characteristics of small data volume and large feature dimension after feature extraction,the LightGBM algorithm with better performance in machine learning is selected to detect the edge contour of point cloud data.In the process of experiment,this paper designs an iterative model,which improves the performance of the classifier through multiple iterations and obtains ideal experimental results.Finally,aiming to solve the problem that the segmented edge contour is wide and contains a lot of noise,this paper designs an optimization algorithm,which performs a series of morphological optimization algorithms such as cavity filling,closed operation,skeleton extraction,and edge contour burr removal.Combine,optimize the edge contour information,and obtain the final scene segmentation result through the connected domain algorithm.By comparing with the traditional algorithm,we can find that our algorithm has better performance and is more robust to depth image noise.