Point cloud segmentation method for a power corridor scene

Abstract

The invention relates to a point cloud segmentation method for a power corridor scene, which comprises the following steps of: training and adjusting a convolutional neural network with a structure with six stages, and transmitting disordered point clouds as input to the convolutional neural network which is trained and adjusted for segmentation. The method has high change adaptability to angle and shape changes of objects in an electric power corridor scene; moreover, the convolutional neural network does not add artificial preset characteristics, and the object is segmented through a learning method, so that the objective facts are reflected better. The work of using the convolutional neural network method is to design a network architecture, which is time-saving and labor-saving compared with the design of manual features. In addition, compared with a deep learning network using voxels, the used convolutional neural network directly uses point cloud input, and the calculated amountand the intermediate variable storage space are greatly reduced. And finally, as there are few data sets of the electric power scene, the data sets are increased by using a transfer learning method, and the generalization ability of the model is improved.

Description

technical field [0001] The present invention relates to the technical field of point cloud segmentation, in particular to a point cloud segmentation method for power corridor scenes. Background technique [0002] A point cloud is a series of unordered points that contain coordinate information and other optional information (such as color information). Point cloud semantic segmentation is to classify each point and assign a label to each point. [0003] Existing solutions: Some scholars use manual features to identify objects, such as using the density of point clouds and the reflection intensity of radar to segment towers and power lines (reference paper: "Rapid Extraction of Power Lines from Airborne LiDAR Data in Space Domain Segmentation_ Liu Zhengjun"). Some scholars convert the point cloud into voxels as the input of the neural network, using the method of deep learning (reference paper: A deep representation for volumetric shapes). [0004] However, there are many ...

AI Technical Summary

Problems solved by technology

[0006] 1. Manual features are relatively fixed, and it is difficult to adapt to changes in the shape of objects

[0007] 2. Manual features reflect people's subjective ideas, which may be different from objective reality

[0008] 3. Manual features require a lot of time and manpower to design

[0009] 4. For the power corridors in different places and the point cloud files scanned

Images