Research On Visual-words Based Object Detection Method

Object detection is an important technology related to computer vision that focuson detecting instances of semantic objects, such as humans, buildings, or cars, indigital images and videos. Well-researched domains include pedestrian detection andcar detection. However, the problem of big data is rarely discussed under this context,most popular methods may hard to deal with the tasks of detecting objects in largescale image sets.Bag-of-visual-words (BoW) image representation has been successfully appliedin many applications of computer vision and multimedia mainly because of itsefficiency in dealing with large scale image set. This method represents local featuresin images as “visual words” and represents images as documents of visual words.Owning to the indexability of visual words, visual words based methods are oftenvery efficient.In this paper, we proposed a novel object detection method based on visual words,to detect object in large scale image set. In our method,“visual-key-words” ofparticular objects is selected form visual words by utilizing a text analysis modelLatent Semantic Analysis (LSA). Obtaining the visual-key-words, the object model,which describes the structure of the object, is constructed. The paper also make adiscuss on overcoming the “synonymy” and “polysemy” problems of visual words indetection task. The proposed method is scale and rotational invariant, and showsrobustness under occlusions and cluttered background. The accuracy and efficiency ofthis method is tested in variety of detection task. In the last of this paper, theapplication and prospects of the method is analyzed.