A new SLAM measurement model based on omni-directional vision and odometer is proposed in this paper. A virtual stereo vision composed of an omni-directional vision sensor and an odometer. Scale Invariant Feature Transform is used to extract stable and available vision features from the omni-images. The 3-D locations of these features are initialized by the pixel coordinates and the odometer data by stereo projection, and the locations will be corrected during the SLAM process when they are observed again. It is demonstrated that the new model can make a good accuracy with FastSLAM algorithm, and the accuracy is greatly improved corresponding to the classical vision sensor.