In this paper, we describe an approach for finding character information in a scene image and estimating the camera viewing direction relative to the found character board. First, character candidates are extracted as subregions with high spatial frequency and large variance of grey-level. Next, we select characters using several heuristics about characters and character arrangements. Furthermore, by calculating the coordinate value of the vanishing point of two parallel signboard edges in a scene image, the relative orientation of the camera to the signboard can be estimated from the single view, which is useful for repositioning the robot and capturing a new image with less distortion of characters. Experimental results indicate the effectiveness of the proposed method.
展开▼
