Research On The Object Recognition Based On SIFT And NDLT

With the improvement of living standard, intelligent domestic robots have extensive applications in cleaning, moving goods, controlling home appliances, answering the phone, and many other home work areas. Geometric distortions(including scaling and rotation) and noise interference exist in the images which are acquired from the intelligent domestic robots. Aimed at the problems, the object recognition algorithm based on SIFT(Scale Invariant Feature Transform) and NDLT(Normalized Direct Linear Transformation) is proposed.In this paper, firstly, the state-of–the-art of feature points extraction and matching which are the most important technologies in object recognition is discussed, and the approaches of extraction and matching are introduced in detail. Secondly, targeted to the existing problem of object recognition of intelligent domestic robots, the advantages and disadvantages of the current feature extraction and matching techniques and traditional DLT algorithm are analyzed. And the overall algorithm design is brought forward. Thirdly, based on analyzing the theory of SIFT matching including the SIFT feature points extraction and matching, the superiority of SIFT matching algorithm against image scaling, rotation and noise are discussed emphatically, and the SIFT feature matching algorithm is used in object recognition. Finally, considering the shortcoming of traditional DLT(Direct Linear Transformation) with low accuracy and the advantage of the SIFT matching algorithm against image scaling, rotation and noise, a kind of object recognition algorithm which unifies SIFT and NDLT is presented. Simultaneously, the demonstrative experiment is given based on MATLAB.The experimental result indicates that the algorithm proposed in this paper, to be used in the object recognition, is both high recognition accuracy and good robustness against various disturbances such as scaling, rotation with a small angle and noise, that meets the requirement of intelligent domestic robots on object recognition. In addition, the users can add target images into the model database according to their needs in order to realize multi-view object recognition, so the algorithm has good flexibility.