Towards Robust Global VINS: Innovative Semantic-Aware and Multi-Level Geometric Constraints Approach for Dynamic Feature Filtering in Urban Environments
Mahamoud Adham, Wu Chen, Yaxin Li, Tianxia Liu · IEEE Transactions on Intelligent Vehicles · 2024
In real-world scenarios with predominant dynamic objects, achieving robust and accurate positioning using Visual-inertial navigation systems (VINS) poses a challenge because these objects dislocate visual features, resulting in degraded feature tracking accuracy, pose deviations, and trajectory drift. Thus, static scene assumption, as proposed in some existing studies, fails in such scenarios. Meanwhile, directly removing potential dynamic objects (either stationary or moving) using deep learning methods degrades accuracy in low-texture scenes, while using image geometric constraints poses challenges when moving objects dominate the scene. To address this, we introduce a real-time global VINS that incorporates an innovative semantic-aware and multi-level geometric constraint approach for better handling of moving objects. Precisely, a feature grading module combining the power of scene cognition and spatial feature extraction is developed to categorize tracked features. This module integrates multi-geometric constraints and semantic information to effectively eliminate dynamic features and employs VI-based motion consistent constraint to eliminate missed detection of the moving objects. Then, a descriptor-matching tracker module is applied to reject mismatches and enhance feature-matching reliability. Uneven feature distribution issue resulting from intensive dynamic feature elimination is addressed by a proposed geometry feature distribution-based auto-adaptive covariance estimation Algorithm. The backend handles long-term pose estimation drift by developing an adaptive multilayer VI-GNSS optimization framework that integrates a subsystem failure detection mechanism. The system performance demonstrates efficient identification and exclusion of dynamic features while retaining static ones. Experimental validation conducted on various datasets in urban dynamic areas reflects the superiority of the proposed method in accuracy and robustness.