Point-State Segmentation for Dynamic Scenes via Probability Model and Orientation Consistency

In this article, an innovative simultaneous localization and mapping system for monocular vision in dynamic scenes is proposed, which eliminates the unreliable features via the point-state segmentation strategy. A neighbor-based random sample consensus method and image-sequences-based selection method are presented to ensure a stable and precise initial state of the system. Based on the accurate initial information, the probability model and orientation consistency method are proposed to degrade the disturbance of the dynamic environment factors in the process of feature tracking. Additionally, a hybrid robust weight model is performed to reduce the influence of dynamic residual terms on trajectory estimation. A series of ablation experiments are designed to fully verify the effectiveness of all proposed innovations. Moreover, the significant performance of the proposed method is demonstrated by comparing it with other state-of-the-art methods on two public datasets and real-world scenes. The experimental results show that the proposed method can effectively eliminate the impact of unstable landmarks for pose estimation. More importantly, experiments also illustrate that the proposed method can achieve a more accurate and robust trajectory than the deep learning-based methods in some challenging scenarios.