In the simultaneous localization and mapping method of robots， the visual scheme has poor stability in indoor scenes with insufficient texture， and the use of structured assumptions can alleviate the above problems. However， if the indoor scene does not strictly meet the structured assumptions， it will lead to greater pose drift. More general structural assumptions and more reliable loopback detection methods can help solve the above problems and improve the robustness of indoor scene visual positioning. To this end， this paper proposes a robust multi-layer spatial structure assumption visual SLAM method. This method makes full use of the structured information in the scene， uses the main direction constraint to define the scene to assist the positioning， and uses a lightweight structured hypothesis loopback to reduce the cumulative drift， so as to construct a high robustness and low drift simultaneous localization mapping algorithm. We conduct a large number of experiments on real vehicle data and open source data sets. The experimental results show that the proposed method has higher positioning robustness and accuracy performance than other open source methods. The loopback detection method has a higher detection rate， and the positioning accuracy in the closed-loop scene is improved by 31.8% on average.