This paper presents a robust Visual-Inertial Odometry that leverages point and structural line features for complex environments with lighting changes, low textures or repetitive textures. Firstly, a front-end with structural line features is designed. Based on the Atlanta World hypothesis, structural line parameterization, extraction, and triangulation methods are proposed for lines with vertical and horizontal direction; Secondly, a non-linear optimization back-end based on sliding-window is carried out, which is tightly coupled with structural line residual. Finally, the localization accuracy of the odometry is verified through real-world experiments.