Feature matching and object registration based on convex hull alignment
Xueyang Xu, Zhiyong Zhang, Changzhen Qiu, Ziwei Liu · 2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA) · 2022
Feature matching on target objects has always been a fundamental problem of the field of computer vision. It plays an important role in applications such as autonomous driving, 3D reconstruction, target recognition, and object registration can be formulated as matching image features to model features. Common feature matching algorithms like SIFT and FAST are mostly used in recent years. By searching for candidate feature points in the whole image, predefined descriptors are calculated for these feature points, and finally matching is performed between different images. Although over the decades, various improvements have made such methods more efficient and stable, and have proven successful in numbers of applications, they are still unstable in the face of occlusion and complex background. In complex scenarios, such methods always encounter problems such as time-consuming and multiple inaccurate matching. In this paper, we proposed an alternative approach to feature matching and object registration by aligning the convex hulls of target objects detected in images. At the same time, this method uses a descriptor based on the convex hull, which can guarantee rotation invariance and partial scale invariance. this alignment method uses a method derived from Convex Template Instance descriptor, which was originally proposed to fruit recognition and classification of segmented objects. We adopt this approach and improve it for the feature matching problem. The experimental result shows that this method can match the feature points on the 3D target quickly and accurately, also perform good robustness.