To solve the problem that the accuracy of visual odometry was decreased in the weak texture environment, an optimization algorithm of binocular visual odometry is proposed. Firstly, the planar constraints are provided for the features without stereo matching by extracting the local feature plane, to increase the quantity of effective stereo features. Secondly, in feature tracking, the uniform acceleration motion model is used during feature tracking to increase the quantity and improve the quality of tracked features. Finally, in pose estimation and optimization, the bundle adjustment method which considers the feature confidence is adopted to reduce the influence of distant feature, and improve the accuracy and robustness of the algorithm. The experimental results based on the datasets and actual scene show that the proposed algorithm has obvious optimization effect on the positioning accuracy under weak texture environment with little computational resources, and has adaptability in other scenes.