Features Extraction And Classification Of Visual-map Based On Bag Of Words

In recent years, with a significant growth in both popularization and function of mobile phones, diverse requirements are gradually proposed by the subscribers of smart phones. Due to the large scale of public buildings and the limitation of position-identifing indications, it is hard for subscribers to obtain map information such as their current location. Thereby, location based services under indoor invironment, namely indoor localization, has become a heating requirement of users’ making use of their smart phones. Visual localization technology utilizing image information has appeared to be a popular research topic in the field of indoor localization, for its easy operation of capturing videos and images, as well as there is no need to deploy any infrastructure in advance.In order to quickly retrieve an image which is a match with the user input image from a huge image database, so that it is made to be the basis for rapid visual indoor localization, the method of fast image retrieval and matching is crucial. To solve this problem, a feature extraction and classification algorithm based on bag of visual words is proposed in this subject, which enables fast retrieval to the database images and location information in visual localization.Firstly, purpose and significance of this research and research status of related fields and topics are elaborated, including bag of visual words model, image features extraction and supervised classification algorithm.Secondly, four mature image features extraction algorithms, such as SURF, BRISK, perceptual hashing and gist algorithm are researched and implemented respectively. The performance as well as advantages and disadvantages of these four algorithms in actual image matching process are analyzed in the subject. As a result, gist features extraction algorithm is chosed as the basic of the following study.Moreover, focusing on the unsatisfied performance of image retrieval and matching utilizing only global feature descriptors, bag of visual words method using global features of images is proposed. This method is based on improved gist features extraction, using K-means clustering algorithm to classify the extracted features, so that visual words are obtained. Thus, by comparing the histograms of visual words, fast retrieval and matching of images is archieved, and this method turns out to be with a satisfied performance.Finally, considering the requirements of image retrieval and matching speed and accuracy during the online stage, supervised images classification algorithm based on bag of visual words is proposed. In this algorithm, an image classifier is obtained through a supervised learning procedure during the offline stage, then user input images are classified using this classifier during the online stage. Analyses of the proposed features extraction and classification algorithm based on bag of visual words indicate that, for image database captured from actual localization environment, this algorithm is able to achieve fast retrieval and matching of user input images, and the process of instant visual localization is improved.