In order to improve the accuracy of the stereo visual odometry in complex environment, a method of stereo visiual odometry with multi-pose estimation constraints is proposed. Firstly, the mathematical models of matching points with the known depth and the unknown depth are established respectively, and then an improved 2D-2D pose estimation model considering the depth of unknown points is proposed, so that the image information can be used fully. Then, the 3D-2D pose estimation model is improved based on the keyframe mappoints, and the keyframe mappoints are updated according to the mappoints corresponding to the current frame. Therefore, the number of matched features can be more and the accuracy can be increased. Finally, according to the improved 2D-2D and 3D-2D pose estimation model, the pose estimation model with multi-pose estimation constraints is eatablished. Then the local bundle adjustment method is applied to optimize the estimated pose, so that the accuracy will be high and the cumulative error will be small. The experiments based on the datasets and the online experiment based on the actual scene show that, this method not only can meet the requirements of real-time location, but also can improve the accuracy of the mobile robot autonomous localization effectively.