Estimation of Rotation of Multiple-View Camera with Variation of Feature Sampling Region

This study proposes a method for estimating rotation of camera based on homography estimation from multi-view images captured by a three-view thermal camera. To address the difficulty in extracting correspondences due to the minimal overlapping regions among the three views, each viewpoint is independently rotated to increase the number of correspondences. Additionally, DoG (Difference of Gaussian) filter was applied to enhance the edge detection performance of thermal images, and the effects of varying sampling region selection methods prior to homography estimation results is analyzed. A comparison of four sampling region selection methods reveals that the most accurate geometric alignment and rotation matrix is obtained when correspondences were selected from widely distributed regions. This finding demonstrates that the method of selecting correspondences significantly influences the accuracy of homography and rotation estimation. The proposed method of rotation estimation based on homography estimation through optimized sampling region selection is validated through experiments using thermal cameras. To substantiate the proposed method, image blending is carried out to validate estimation accuracy from visual perspective.