RVP-VIO: a refined and robust vanishing point (RVP)-assisted visual inertial odometer (VIO) in a structured environment

Abstract Current visual-inertial odometry (VIO) still suffers from error drift due to factors such as lighting variations and low-texture features. Vanishing points (VP), as potential global information in structured environments, can explicitly restore the attitude relationship between the platform coordinate system and the world coordinate system. VIO enhanced with VP (VP-VIO) has been proven to be an effective augmentation technique. However, existing VP extraction algorithms still face the issue of ‘pseudo VP (PVP)’, and the use of empirical stochastic models during the attitude constraint measurement update phase further limits the enhancement performance of VP-VIO. Therefore,in this paper, we propose a refined VP (RVP)-assisted VIO algorithm. Specifically, for candidate VPs extracted through preliminary resampling, an optimal VP extraction method based on the median is proposed, leveraging inertial prior constraints and considering the orthogonal properties between VP vectors. Additionally, during the filter measurement update phase, the IGGⅢ variance inflation model from the field of geodesy is introduced to further mitigate the negative impact of abnormal VP on attitude constraints. In the experimental section, the proposed RVP is first validated using a public dataset, showing higher accuracy compared to traditional methods, with an improvement of approximately 12.9%. Regarding VIO positioning performance, due to the higher accuracy of RVP and the measurement update model based on adaptive noise, the proposed RVP-VIO exceeds SOTA performance in mixed Manhattan world compared to several representative VIO methods.