Efficient Vanishing Point Estimation for Accurate Camera Rotation Estimation in Indoor Environments

Point-based simultaneous localization and mapping (SLAM) algorithms tend to degrade in texture-less indoor environments. Considering the pervasive presence and structural regularity of line features in artificial scenes, many researchers have developed line-based SLAM systems incorporating structural information, such as vanishing points (VPs). However, VP estimation introduces additional computational resource consumption, which degrades the real-time performance of SLAM systems. In this letter, we propose an efficient VP estimation algorithm that is applicable under both the Manhattan World (MW) and Atlanta World (AW) assumptions. Based on the assumption that each set of three orthogonal VPs is associated with a local Manhattan frame (LMF), the algorithm exhaustively searches all the existing VPs and models their associated LMFs. This process subsequently allows for absolute rotation estimation between the camera and fixed LMFs. Furthermore, we propose a translation-constrained line optical flow tracking (TC-LOFT) model to accurately and efficiently track the structure line feature. By building the residual of prior VP estimation and tracked structure line features, we propose a progressive absolute rotation refinement (PARR) to update VP estimation across consecutive frames. Experimental results on public datasets demonstrate that the proposed algorithm achieves significant efficiency, with negligible runtime for VP estimation, and high accuracy with less than 5 degrees of rotation estimation error in general indoor environments. The proposed algorithm outperforms the state-of-the-art and addresses the gap in efficient VP estimation under AW assumption.