Data clustering-based robust scale invariant feature transform (SIFT) feature matching method

Abstract

The invention discloses a data clustering-based robust scale invariant feature transform (SIFT) feature matching method, which comprises the following steps of: (1) acquiring reference image sequences, extracting SIFT feature sets, clustering all the SIFT feature sets by adopting a synthetic k-d data structure, and merging the repeated feature sets as clustering features, wherein the feature descriptors of the clustering features are expressed by adopting repeated feature average descriptors; (2) acquiring real-time image sequence feature sets by using an SIFT method, performing robust matching on the reference image sequences and the real-time image sequences, and selecting the corresponding reference image with maximum feature points as a key image to complete feature matching of a first stage; and (3) completing feature matching of a second stage by using the key image, removing exterior points by adopting the technologies of random sampling consistency (RANSAC), basic matrix and the like, and finally merging the feature matching results of the two stages. According to the matching method, the noise information interference can be reduced, and the robustness of feature matchingis greatly improved.

Description

technical field

[0001] The invention belongs to the technical field of computer augmented reality, and specifically performs robust matching on a SIFT feature set under the condition of a large amount of data, eliminates the influence of noise in the matching process, and improves the robustness of feature matching. Background technique

[0002] SIFT feature matching technology is an important research content in computer image and vision, and has been widely used in many fields such as image retrieval, target recognition, 3D reconstruction and camera pose recovery.

[0003] By treating SIFT feature descriptors as Cartesian coordinates in high-dimensional Euclidean space and using Euclidean distance as a similarity measure, the SIFT feature matching problem can be transformed into a query problem of the nearest geometric points in Euclidean space. As one of the most important branch and bound methods, the k-d data structure can effectively realize spatial division and is wid...

Examples

Embodiment Construction

[0020] Table 1 has provided concrete steps of the present invention:

[0021]

[0022]

[0023] Table 1

[0024] 1. Synthetic k-d data structure construction and feature clustering in the offline stage

[0025] figure 1 A schematic diagram of the synthetic k-d data structure is given.

[0026] In the offline stage, the reference image SIFT feature set is obtained first, and all reference image SIFT feature sets are clustered. In the feature clustering method, common features are obtained directly from the image by the SIFT algorithm, which can be denoted as F s ={b i , d d}, where b i is the subordinate image, d d is a feature descriptor. The clustering feature is a collection of common features, denoted as F c ={F ss , d md}, where F ss is the general feature set, d md is the cluster average descriptor, its value is: where k is the number of common features included in the clustering feature, (d d ) i is the i-th common feature descriptor. This value i...

technical field

[0001] The invention belongs to the technical field of computer augmented reality, and specifically performs robust matching on a SIFT feature set under the condition of a large amount of data, eliminates the influence of noise in the matching process, and improves the robustness of feature matching. Background technique

[0002] SIFT feature matching technology is an important research content in computer image and vision, and has been widely used in many fields such as image retrieval, target recognition, 3D reconstruction and camera pose recovery.

[0003] By treating SIFT feature descriptors as Cartesian coordinates in high-dimensional Euclidean space and using Euclidean distance as a similarity measure, the SIFT feature matching problem can be transformed into a query problem of the nearest geometric points in Euclidean space. As one of the most important branch and bound methods, the k-d data structure can effectively realize spatial division and is wid...

Examples

Embodiment Construction

[0020] Table 1 has provided concrete steps of the present invention:

[0021]

[0022]

[0023] Table 1

[0024] 1. Synthetic k-d data structure construction and feature clustering in the offline stage

[0025] figure 1 A schematic diagram of the synthetic k-d data structure is given.

[0026] In the offline stage, the reference image SIFT feature set is obtained first, and all reference image SIFT feature sets are clustered. In the feature clustering method, common features are obtained directly from the image by the SIFT algorithm, which can be denoted as F s ={b i , d d}, where b i is the subordinate image, d d is a feature descriptor. The clustering feature is a collection of common features, denoted as F c ={F ss , d md}, where F ss is the general feature set, d md is the cluster average descriptor, its value is: where k is the number of common features included in the clustering feature, (d d ) i is the i-th common feature descriptor. This value i...

Images