Visual simultaneous localization and mapping(SLAM) plays an important role in the outdoor work of robots, such as field exploration, positioning and reconnaissance. In order to make the robots work better outdoors, this paper proposes a fully online real-time binocular direcnt vision SLAM algorithm which is not limited by the fixed vocabulary of the bag of word. As a direct vision SLAM, the system mentioned in this paper can make use of any image pixel with sufficient intensity gradient to make it still have strong robustness in the area lacking feature points. The system algorithm introduces binocular static residual constraint and sliding window optimization to remove occlusion to enhance the tracking accuracy of the system, adds closed-loop detection and pose map optimization modules, and establishes an online bag of word, so that the system can still work in a large-scale and unfamiliar environment. The performance of this algorithm is evaluated on the public EuRoC dataset and KITTI dataset. The positioning accuracy of the proposed system is better than the most advanced direct visual SLAM system, and is robust to indoor scenes and outdoor scenes.