Abstract: In this paper,we present a method to extend commonly used two-frame rotary motion estimation techniques to multi-frame estimation by exploiting multi-frame subspace constraints.At first,we show that the image motion from a rotating camera across multiple images is embedded in a lower dimensional linear space if the camera’s intrinsic parameters keep unchanged；then a SVD technique is used to project columns of the motion matrix into a lower dimensional linear space；and finally,a least squares method is used to estimate the motion parameters.The proposed method needs neither prior scene 3D information,nor camera’s intrinsic parameters,is shown capable of achieving more accurate alignment results as multiple frames usually have more constraints than only two frames.The method can also be used for the rotary motion estimation in a small image region.