Image feature extraction and matching is a key problem in many applications of computer vision, such as target detection, object recognition, 3D reconstruction, image registration and video understanding. Due to a large family of camera lens types, various image content, and diversity of applications,image feature extraction and matching has always been a challenging problem,and a lot of researchers are actively engaged in the field. In this dissertation, the main work is focused on applications of feature extraction and matching in some real problems. The main points are as follow: ² An approach for matching edge points between two fish-eye images is proposed. In Lowe's original scheme, those points along edges are eliminated during the feature extraction process since the difference-of-Gaussian function's response is unstable to even small amounts of noise, whereas correspondence along edges play a very important role in most applications as 3D reconstruction, object detection, image registration, etc. We attempt to tackle this problem by: Firstly, MSER regions are extracted and matched using SIFT descriptor and epipolar line constraint. Within correspondent regions, intersection of epipolar line and region edges constrain the possible point correspondences, and finally, homagraphy is invoked to reject possible mismatches. Our preliminary results on fish-eye images show that the obtained edge points are viable for 3D reconstruction. ² A new automatic wheel contour detecting approach is proposed. As the contour of wheel rim is easier to detect than the contour of wheel, and the contour of wheel rim and the contour of wheel are mostly concentric circles on the same plane, we first detect the contour of wheel rim, the center of wheel rim and the contacting point between the wheel and the ground, then the contour of wheel is detected via the invariance of cross ratio under projective transformation. Our extensive experimental results with real images under various weather conditions show that the proposed approach is capable of detecting contours of wheels robustly and automatically.
修改评论