The invention discloses a face alignment method based on cascade position regression of random forests. The face alignment method is characterized in that 1, a normalized face image is acquired; 2, the average shape of the faces can be calculated; 3, a candidate feature point of a face alignment frame is generated; 4, a face shape index grey level is generated; 5, a face shape index feature X is generated; 6, a face alignment frame is built; 7, the face shape is initialized, and after the continuous iteration, the final estimated face shape can be output. The good robustness can be kept under the conditions of the changing illumination, the changing expressions, and the shielding, and the precision can be improved, and the failure rate can be reduced.