Semantic point cloud acquisition and semantic octree map construction method based on score-to-score

Abstract

The invention provides a semantic point cloud acquisition and semantic octree map construction method based on division-to-division, which comprises the following steps of: performing semantic acquisition on an acquired image by using a first image semantic segmentation network to obtain a primary segmentation result; sequentially connecting an image restoration network and a second image semantic segmentation network in series behind the first image semantic segmentation network to form a sis processing framework; determining a processing area of the image restoration network, outputting a corresponding result to a second image semantic segmentation network by the image restoration network, and outputting a second segmentation result by the second image semantic segmentation network; according to a priority mechanism, optimizing and complementing to obtain a final optimized semantic segmentation result; images and depth information are obtained through a camera, and dense point clouds are obtained in combination with SLAM system pose information; in combination with depth information, an optimized image semantic segmentation result is mapped to dense point clouds to obtain dense point clouds with semantics, and an octree semantic map with more complete and more stable semantic information can be obtained.

Description

technical field [0001] The invention belongs to the fields of artificial intelligence robots and computer vision, and in particular relates to a method for acquiring semantic point clouds of images and constructing semantic octree maps based on a "segmentation-repair-segmentation" framework. Background technique [0002] Most artificial intelligence robots construct semantic maps by acquiring image information through visual simultaneous localization and mapping (SLAM) systems. The visual SLAM system filters out key frames for pose estimation at the front end. These key frames combine with its image semantic segmentation information and pose depth and other information to act on the mapping module to construct a semantic octree map for navigation. At present, visual SLAM systems mostly use a single semantic segmentation method to obtain image semantic information, however, there are a large number of dynamic objects in real scenes, such as walking people, moving cars, etc. ...

AI Technical Summary

Problems solved by technology

[0003] The present invention aims at the missing or wrong semantic information of the octree semantic map constructed by the visual SLAM system in a dynamic scene, and provides a network framework SIS for optimizing the image semantic segmentation results of key frames when the visual SLAM system acquires semantic information , to improve the semantic map information

Images