Estimation of Camera Rotation Using Quasi Moment Features
Hiroyuki Shimai, 敏勝 川本, Takaomi Shigehara, Taketoshi Mishima, Masaru Tanaka, Takio Kurita · IEICE Transactions on Fundamentals of Electronics Communications and Computer Sciences · 2000
SUMMARY We present two estimation methods for camerarotation from two images obtained by the active camera beforeand after rotation. Based on the representation of the projectedrotation group, quasi moment features are constructed. Camerarotation can be estimated by applying the singular value decom-position (SVD) or Newton’s method to tensor quasi moment fea-tures. In both cases, we can estimate 3D rotation of the activecamera from only two projected images. We also give some ex-periments for the estimation of the actual active camera rotationto show the effectiveness of these methods. key words: quasimomentfeature,projectedrotationgroup,sphericalharmonics,activecamera 1. IntroductionRecently the active camera has been popular in variousfields, because it is getting lower in cost and higher inits performance. It is quite important to estimate thecamera rotation from two images obtained by itself be-fore and after rotation for the self-calibration problemof the active camera and the control of a robot withvision system, etc.Usually the estimation method for the camera ro-tation is based on the point matching [8]and/or mo-ment features [9]. However, the point matching methodhas two problems, one is the difficulty of selecting thecollateral points and the other is the unstableness un-der noises. The method based on moment featuresdoesn’t have these problems because they are the in-tegral quantities. However the moment features havealso two problems. One is that marginal regions of thescreen contribute dominantly rather than the center ofthe screen. It causes the loss of the important imageinformation because the target image is usually locatedat the center of the screen. The other is that, it ismore serious, moment features can only estimate therotation around Z axis along the light axis of the lens.Thus, they cannot estimate the rotation included panand/or tilt, this is because pan and tilt can be rep-resented by 3D rotations. Ordinally, moment featuresdon’t take the projection from 3D space onto 2D planeinto consideration.