Three-dimensional point cloud semantic segmentation method and system based on dynamic aggregation

Abstract

The invention relates to a three-dimensional point cloud semantic segmentation method based on dynamic aggregation, and the method comprises the steps: constructing a three-dimensional point cloud semantic segmentation model of an encoder composed of a sampling module and a cascaded dynamic aggregation module, and a decoder composed of a feature propagation module; and performing feature dynamic aggregation on the point cloud data output by the sampling module through a dynamic aggregation module to output a feature map, transmitting the feature map to a full connection layer through a feature propagation module, and outputting a category. Compared with the prior art, the method has the advantage of high semantic segmentation precision.

Description

technical field [0001] The invention relates to the field of image processing, in particular to a dynamic aggregation-based three-dimensional point cloud semantic segmentation method and system. Background technique [0002] In recent years, with the development of depth sensor hardware and deep learning technology, 3D point cloud semantic segmentation technology based on deep learning has gradually become a research hotspot, and is widely used in autonomous driving, 3D scene reconstruction, robots, etc. [0003] 3D point cloud data is usually a collection of a series of points obtained by 3D scanning equipment such as lidar, which contains 3D spatial information (X, Y, Z coordinates in space), and may also have object information (such as color, intensity value) , reflection value and other characteristics. The representation of 3D point cloud data is simple, compact, and rich in information, so it is widely adopted as mainstream 3D data in 3D deep learning. However, 3D p...

AI Technical Summary

Problems solved by technology

[0005] 1) The weight sharing of multi-layer perceptron leads to limited feature extraction ability, which limits the ability of the model to capture spatial semantic information in large-scale scenes

[0006] 2) The 3D point cloud data is more sparse than the 2D image, and the original 3D point cloud data is directly trained, and the weight of its semantic features changes unstable, which affects the model performance;

[0007] 3) In the encoder, the aggregation operation usually selects the representative features of each local area by selecting the non-learned maximum value, average value, etc. The semantic feature capture ability in the local area is insufficient, and there is still huge room for performance improvement

Images