A 3D Point Cloud Scene Segmentation Method Based on Instance Embedding and Semantic Fusion

Abstract

The invention discloses a three-dimensional point cloud scene segmentation method based on instance embedding and semantic fusion, belonging to the technical field of image processing. Aiming at the technical problems existing in the 3D scene segmentation method based on deep learning, the present invention proposes a 3D point cloud scene segmentation method based on instance embedding and semantic fusion. Firstly, point cloud mapping and feature expansion are carried out, and then deep neural network model setting and training are carried out, and the segmentation processing of 3D scenes is realized based on the trained neural network. At the same time, a new CRF is also introduced to optimize the segmentation results. The present invention is based on the geometric characteristics of the 3D point cloud. Except for the basic information of the point cloud, the extracted feature vectors extend a series of new attributes to describe the 3D point through the local and neighborhood attributes of each point of the 3D point cloud. The characteristics of the cloud: scattering, linearity, planarity and verticality, improve the segmentation performance; and further improve the segmentation performance of the 3D point cloud scene based on the new CRF function model introduced.

Description

Technical field [0001] The present invention belongs to the field of image processing, and more particularly to three-dimensional field segmentation techniques that combine the semantic splitting instance segmentation and semantic segmentation. Background technique [0002] With the increasing popularity of low-cost 3D sensors (such as Kinect) and light field cameras, there are many 3D applications, such as automatic driving, robots, mobile navigation, virtual reality and 3D games. Automatic understanding of 3D dot clouds, segmentation recognition has become an important research branch of computer graphics, and many research are committed to this field and have achieved many results. For example, in automatic driving, three-dimensional dot cloud segmentation technology provides many real-time aids for the processing, understanding, identification of the scene, helping the driving system faster and accurate planning path schemes and grasping bursts. [0003] 3D data and two-dimen...

AI Technical Summary

Problems solved by technology

But the first problem with this method is how to choose the angle of each view in the view set

The second problem is that the view set does not fully express the 3D model. It only expresses the data of each perspective, and a large amount of positional space data is lost, which causes a large error in the segmentation of the 3D scene, resulting in unsatisfactory segmentation results.

If the grid side length is too large, more precision will be lost, and at the same time, it will cause very low quantization noise errors, which will limit the accuracy of the segmentation results. On the contrary, if the grid side length is too small, many blank areas will be generated, resulting in a waste of computing resources.

And the time and space complexity of this method will increase to O(n 3 )increase

Images