Although motion recovery for limited field-of-view (FOV) cameras has been studied for decades under the pin-hole projection, few attentions have been paid on omni-directional cameras. Omni-directional cameras are usually constituted complicatedly, especially some are made of several limited FOV cameras, and thus motion recovery for omni-directional cameras is more different. In this paper, an algorithm is presented to estimate motion of the omni-directional camera under the spherical projection.  Unlike motion recovery algorithms for traditional limited FOV cameras, in which the intrinsic matrix plays an important part, this algorithm only estimates extrinsic matrices without the influence of intrinsic matrix. So it is more efficient and robust. To reduce computation, a method is also presented to select key frames using the translation offset of neighboring frames and remove some mismatched feature points. The experiments verify the efficiency of the algorithm and the visual result is demonstrated through a simulation experiment.

Figure：Simulative demonstrations of the panoramic spheres and the 3D points