A Fast And Robust Building Recognition Algorithm And System

As an important part of location recognition, building recognition’s research work issignificant. This paper focuses on a particular scene--building recognition and carries outresearch work aiming to propose a fast and robust building recognition method and establishthe appropriate system.To achieve this goal, we firstly summarized several types of main algorithm in the area ofimage target recognition. By analyzing their advantages and disadvantages, we selected thelocal invariant features to handle the building recognition problem. Then we studied theprinciples and extraction process of the classical local invariant features–SIFT, and carriedout some building recognition experiments using SIFT features. The experimental resultsshowed that the SIFT features had good validity in terms of building recognition, but theirextracting time was so long that they could not be used in a practical application. As a result,we did a depth research on the improved algorithm of SIFT–SURF. It was found that theSURF features not only retained the recognition effectiveness of SIFT features, but also had avery short feature extraction time, which was suitable for building recognition.However, if aimed to create a fast and robust building recognition system, we had toimprove the SURF algorithm on the following three issues: not making full use of the image’scolor information, low feature matching efficiency and feature mismatch. Consequently, wefirstly used the color quantization histogram in the HSV color space as the color feature tofilter the database images, only retaining images with similar color of the query image, whichcould significantly reduce the image number of follow-up match. Secondly, we introduced theK-D tree algorithm to index all the feature points extracted from the database images in thetree structure. After that, we used the K-D tree search algorithm based on BBF to find nearestneighbor and second nearest neighbor of the query feature point in the built K-D tree, whichimproved the image matching efficiency. Thirdly, we combined RANSAC algorithm to do theconsistency test on the successfully matched feature points. The algorithm randomly created anumber of mathematical models and calculated the corresponding numbers of interior points.Then it selected the optimal mathematical model with the largest interior point’s number andused it to remove the error match between images, which improved the accuracy of featurematching. Eventually, we formed a fast and robust building recognition method.Based on the proposed building recognition method, we constructed a fast and robustbuilding recognition system. The system consisted of two modules of the offline processingand online processing, which were responsible for processing the image database and query image recognition request respectively. Subsequently, we used parts of images to do thevalidation experiments on the established system. The results showed that the constructedbuilding recognition system fitted all the technical requirements with high recognitionaccuracy and running speed, which was a fast and robust system.