Visual SLAM method based on semantic segmentation dynamic points

Abstract

The invention discloses a visual SLAM method based on semantic segmentation dynamic points, and relates to the technical field of computer vision. The method comprises the following steps: acquiring environment image information through an RGB-D camera, and performing feature extraction and semantic segmentation on an obtained image to obtain an extracted ORB feature point and a semantic segmentation result; using a dynamic object detection algorithm based on multi-view geometric constraints to detect residual dynamic objects and reject dynamic feature points; and tracking, local mapping and loopback detection threads are executed in sequence, so that an accurate static scene octree three-dimensional semantic map is constructed in a dynamic scene, and finally the visual SLAM method based on semantic segmentation dynamic points facing the dynamic scene is realized.

Description

Technical field: [0001] The present invention relates to the technical field of computer vision, and more specifically, relates to a visual SLAM method based on semantically segmented dynamic points. Background technique: [0002] Simultaneous localization and mapping (SLAM, simultaneous localization and mapping) research has a long history. It was first proposed by Smith et al., and then gradually improved by scholars. It refers to the fact that the robot estimates its own position through the information obtained by the mounted sensors when the environment is unknown, and at the same time constructs a map of the perceived surrounding environment. Visual SLAM is a system that uses cameras as sensors to complete positioning and mapping tasks. It is a prerequisite for mobile robots to complete intelligent tasks, and has become a hot spot in the current research on autonomous mobile navigation of robots. [0003] At present, researchers have found many mature algorithms, such...

AI Technical Summary

Problems solved by technology

However, most visual SLAM schemes assume that the object is still. However, dynamic targets often appear in real scenes, which reduces the robustness of visual odometry, causes inaccurate camera pose estimation, and cannot build a globally consistent map.

In addition, the geometric map constructed by visual SLAM maintains the geometric information in the environment, which is generally a sparse landmark map including spatial feature points, or a dense point cloud map that can be used for navigation and obstacle avoidance, which cannot meet the high-level requirements for robot-environment interaction. Semantic Task Requirements

Images