This paper proposes a new method for rotation-invariant picture matching using an orientation histogram represented by autoregressive model. An orientation histogram is created through gradient of brightness at each pixel. The orientation histogram intersection (OHI) algorithm is effective for searching rotated patterns in the scene, but it needs time consuming circular shifting operations of a reference orientation histogram. This computational demerit increases when registration accuracy is desired. The proposed method expresses an orientation histogram by an autoregressive model as a closed contour, and it has rotation-invariance as well as OHI and then fast computation can be performed by use of the representation with smaller parameters. Experiments with real pictures show effectiveness of the proposed method.