A city-level point cloud semantic segmentation system and method based on spatial perception

Abstract

The application relates to three-dimensional GIS technology, in particular to a city-level point cloud semantic segmentation system and method based on spatial perception, which comprises the following steps: low-order specific feature coordinates and colors are processed through a plurality of groups of feature extraction operations and down-sampling operations, are mapped into multi-scale high-order abstract features, and deep feature information hidden in input data is mined; feature enhancement is realized by inputting the high-order abstract features into two feature enhancement branch modules for processing, the receptive field of a feature map is expanded, and the ability of a model to identify small target objects is improved; high-order abstract features are processed through a plurality of groups of up-sampling and feature learning, are restored into low-order specific features, and are propagated to each point of point cloud data, so that each point is given a semantic label. The method enhances the construction of a model on local features of point cloud, realizes the expansion of the receptive field of the model without losing feature information, and finally realizes efficient city point cloud semantic segmentation.

Images