In order to improve the robustness and accuracy of the visual odometry in the dynamic scene, a stereo visual odometry based on moving object detection is proposed. Firstly, a scene flow calculation model considering the camera pose is established to represent the motion vector of the objects. Secondly, the method of constructing virtual map points is proposed. On the one hand, the motion object detection can be complied according to the virtual map points and the scene flow, on the other hand, the virtual map points ensure that there are still enough matched point pairs to estimate the pose when the proportion of the moving objects in the image is too large. Finally, the feature points in the current frame will be matched with the local map points and the virtual map points, and according to the matching results, the nonlinear optimization model considering the virtual points is constructed to estimate the camera pose. It can not only ensure that the static map points do not match with the feature points of the motion object, but also avoid the failure of the visual odometry when the effective point pairs are too few. The results of the dataset experiment and the online experiment in the actual scene show that the proposed method improves the robustness and accuracy of the visual odometry in the dynamic scene.