Homography based Visual Bag of Word Model for Scene Matching in Indoor Environments

This paper proposes a data driven approach to perform scene localization in indoor environments. The proposed algorithm named p-BoW is designed to cope with self-repetitive and confusing patterns in indoor environments of any type. The algorithm uses the Visual Bag of Words (BoW) model along with proposed voting scheme to perform scene localization from a database of captured images. In the first phase, a small subset of images closer to the query image is found via standard BoW. In the second phase, verification is performed (if required) to identify the best matched image from this subset against the query image. Normalised term frequency (ntf ) weighting scheme has been found to outperform normalised term frequencyinverse document frequency (ntfidf ) scheme in matching precision. Proposed algorithm makes use of visual BoW based on SIFT features in the first phase and perspective transformation during the verification phase for image matching. The resulting proposed system has been able to perform scene matching efficiently in our indoor environment (having about 35 indoor locations) with an accuracy of more than 91% on different cluster sizes.