58 results about "Vocabulary tree" patented technology

The invention discloses a visual-content-based method for establishing a multi-level semantic map. The visual-content-based method comprises the following steps: gathering images shot by a robot wandering in an environment and labeling the scenes of spots for photography; constructing a hierarchical vocabulary tree; constructing a knowledge topological layer so as to grant knowledge to the knowledge topological layer; constructing a scene topological layer; constructing a spot topological layer. According to the visual-content-based method, a visual sensor is utilized for constructing the multi-level semantic map for a space, and digraph structure is used on the knowledge topological layer for storing and inquiring the knowledge, so that unnecessary operation can be eliminated in a knowledge expression system, and the inserting and inquiring speed is quick; the scene topological layer is utilized for carrying out abstract division on the environment so as to abstractly divide the whole environment into subdomains, so that the image searching space and the path searching space can be reduced; the spot topological layer is utilized for storing specific spot images, the self-positioning can be realized by adopting image searching technology, and the error accumulation problem of self-positioning estimation is solved without maintaining the global world coordinate system.

The invention discloses a data-clustering-based adaptive image SIFT (Scale Invariant Feature Transform) feature matching method, which comprises the following steps of: (1) acquiring a reference image sequence, extracting image SIFT feature sets and clustering all image SIFT feature sets by adopting a k-d (k-dimensional) tree; (2) organizing all the image SIFT feature sets by adopting a cascaded vocabulary tree and carrying out secondary feature clustering on the feature sets contained in vocabulary tree nodes; (3) matching BBF (Best Bin First) purified based on a specific value with a dual-mode clustering feature based on comentropy by using the cascaded vocabulary tree to finish the first-stage feature matching; and (4) finishing the second-stage feature matching by using a key image and combining the feature matching results of the two stages; and finally rejecting exterior points by adopting RANSAC (Random Sample Consensus), basic matrix and other technologies. By using the matching method, the robustness of the feature matching can be greatly improved and the adaptability of the feature matching can be effectively enhanced.

The invention relates to an unmanned aerial vehicle aerial image matching method based on vocabulary tree blocking and clustering. Firstly, images in a scene are quantified by using a vocabulary tree, and hierarchical clustering is established for the centralized mass features of the images, so that rapid similarity screening of a to-be-matched image set and a mass image set is realized, a rapid scene classifying process is realized, and frame-after-frame matching and selection of the to-be-matched image set according to a traditional method are avoided. Secondly, a thumbnail of two frames of obtained images having similarity is established, and rough matching is performed on the images under the thumbnail. Then, image blocking is performed by using a clustering method, which is an effective trial for the thought of rough-to-fine matching. Furthermore, the unmanned aerial vehicle aerial image matching method is provided for the first time specific to the data characteristics of large data size of unmanned aerial vehicle aerial images, high image resolution, low image overlapping ratio and the like, so that the unmanned aerial vehicle aerial image matching accuracy and efficiency are effectively increased. The effectiveness of the method is verified by testing aerial images in a PAMView: Providence Aerial MultiView Dataset database.

A method for multi-subtree-based distributed image training and searching comprises the following steps: step 1, selecting initial cluster central points and clustering by computational nodes, and distributing the clustered new cluster central points to each computational nodes; step 2, training task subtrees by each computational node, and taking a cluster central point as a task subtree growing point. The method for multi-subtree-based distributed image training and searching further comprises step 3, extracting feature points from the image to be searched, and transmitting the feature points to corresponding computational nodes according to cluster central point affiliation; step 4, utilizing task subtree to process and transmitting the results to management nodes by the computational nodes, and summarizing the computational results of computational nodes to acquire image searching results by the management nodes. The method can divide the training task of one vocabulary tree into the training tasks of a plurality of subtrees, enable a plurality of computational nodes to process in parallel so as to accommodate larger image training set for strong expansibility, meanwhile reduce the time cost for process of image training and searching.

The invention discloses a GPU acceleration-based large-scale image vocabulary tree retrieval method and belongs to the field of computer vision. Based on structural features of a GPU platform, a novel vocabulary tree data structure is employed; and traditional SIFT feature points are modularized and organized, so GPU large-scale parallel computing capacity can be fully utilized and large-scale image SIFT feature point mapping efficiency can be greatly improved. After batches of SIFT feature points are mapped into a histogram, a quick and compact method is employed to enhance and improve an original reverse index scheme, so post vision vocabulary peculiarity express, histogram normalization and query image grades can be accelerated to a certain extent. The invention further provides a GPU acceleration-based large-scale image vocabulary tree retrieval system, playing an important acceleration role for image retrieval and having a great effect on a large-scale scene three-dimensional reconstruction algorithm efficiency increase.

The invention discloses an unmarked augmented reality multi-target registration tracking method. The method comprises the following stages: 1) an offline stage of performing hierarchical k-means clustering to obtain a vocabulary tree model, registering corresponding image ids for reverse indexes of all leaf nodes of a vocabulary tree, and finally updating the vocabulary tree to a vocabulary tree with tf-idf according to occurrence frequencies of the leaf nodes of the whole tree and a total to-be-registered image quantity; and 2) an online stage which is a real-time response system and comprises the steps of retrieving most similar images according to a trained vocabulary tree for images input in real time by a camera, calculating out to-be-loaded 3D object initial poses through pose estimation of the camera, tracking motions of targets by adopting a KLT tracking algorithm and finally creating augmented reality scenes of the targets through a rendering thread. The method provides an efficient and reliable solution for unmarked multi-target augmented reality.

The invention discloses an image retrieval method, which is realized on the basis of a vocabulary tree level semantic model. Firstly, the characteristics of SIFT (scale-invariant feature transform) comprising color information of an image are extracted to construct the characteristic vocabulary tree of an image library, and a visual sense vocabulary describing image visual sense information is generated. Secondly, the Bayesian decision theory is utilized to realize the mapping of the visual sense vocabulary into semantic subject information on the basis of the generated visual sense vocabulary, a level semantic model is further constructed, and the semantic image retrieval algorithm based on content is completed on the basis of the model. Thirdly, according to relevant feedback of a user during a retrieval process, a positive image expandable image retrieval library can be added, and the high-level semantic mapping can be revised at the same time. Experimental results show that the retrieval method is stable in performance, and the retrieval effect is obviously promoted along with the increasing of feedback times.

The invention discloses a three-dimensional object retrieval method based on multi-view classification. The method comprises the steps of obtaining an original image; processing the original image toobtain a multi-view image; processing the multi-view image to obtain a feature descriptor; processing the feature descriptor to obtain a vocabulary tree; and retrieving a to-be-retrieved image in thevocabulary tree to obtain the serial number of the original image. According to the method, a multi-view classification algorithm is adopted to obtain an omnibearing multi-view feature image of the three-dimensional object in a two-dimensional plane, the three-dimensional object can be directly retrieved, the method is simple, convenient and easy to implement, and the problem that the retrieval performance is reduced due to view angle changes of the object shot by a user in three-dimensional image retrieval is solved.

A method and apparatus for recognizing an object may obtain, from an input image, feature points and descriptors corresponding to the feature points, determine indices of the feature points based on the descriptors, estimate a density distribution of feature points for each of the indices, and recognize an object included in the input image based on the estimated density distribution.