This paper proposes an object recognition method that detects 3D position of objects in real time and achieves adequate accuracy from video sequences captured by a camera. Automatic object recognition from video sequence is a challenging task since it is hard to achieve robustness and accuracy under various scenes. In our proposed method, the object information is supposed to be obtained by media other than captured video such as RFID-tags. The object information includes a 3D model and other features which are effective for video based object recognition. In our proposal, there are four features to achieve robustness and accuracy: i)The proposed method extracts object location in the video using its color information. Moreover, we employ HSV color space instead of RGB color space to cope with luminance variety. ii)To reduce computational complexity, the specification of object position is done only in an initial captured frame and executes matching operation in subsequent frames with limited search region. iii)To limit the range of translation vector, we use the size of a limited search region and the object size. vi)In the final step of matching operation between wire frame model and the detected region, aspect ratio of the extracted object region is utilized to narrow down the wire frame model projected onto images. The performance of the proposed method is improved by the search window which sets on neighborhood of corner point in the projected image.
展开▼
