Visual semantic database construction and global positioning method based on deep learning

Abstract

The invention discloses a visual semantic database construction and global positioning method and system based on deep learning. The method includes: a visual semantic database construction step, obtaining a key frame of a camera and a corresponding pose, inputting a key frame image, obtaining object semantic information of the key frame, extracting object local features according to the object semantic information, and storing the pose, the object semantic information, and the object local features corresponding to the key frame in a database; and a camera positioning step, screening a candidate image similar to a current image by employing a two-layer candidate frame retrieval mechanism including article category coarse screening and image feature fine screening. According to the methodand system, deep learning and a conventional SLAM algorithm are combined, identification and image segmentation of objects in scenes can be effectively realized, and a corresponding visual semantic database is established; besides, the semantic database performs global positioning by employing the two-layer screening mechanism including the object semantic information and the object local features, the matching efficiency can be improved, mismatching is reduced, and the positioning precision is improved.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a deep learning-based visual semantic library construction and global positioning method. Background technique [0002] Traditional object recognition and segmentation methods are usually based on feature point matching to identify different objects, and combine clustering algorithms to complete the segmentation between objects. However, the method based on artificial feature point matching cannot identify objects well in special environments such as illumination and sparse feature points; when the distance between objects is too close, traditional clustering algorithms cannot accurately segment objects. [0003] Compared with the traditional identification and segmentation methods, the object identification and segmentation method based on deep learning is more robust, and it can also accurately identify and segment objects in complex environments such as lighting and ...

AI Technical Summary

Problems solved by technology

However, the BoW bag-of-words model usually uses artificially constructed features and combines clustering algorithms to construct dictionary representation images, and uses dictionary histograms for image matching. It is still prone to mis-matching in complex environments such as lighting and occlusion.

Images