TDHVIO: A Texture Richness Guided Double Hybrid Visual-Inertial Odometry

This paper proposes a texture richness guided double hybrid visual-inertial odometry (TDHVIO), which improves the accuracy and robustness of localization under low texture conditions. Different from existing hybrid VIO frameworks, the proposed framework achieves a tightly coupled double hybrid visual constraints based on the judgment of texture richness, which reduces the both uncertainty of direct and indirect visual joint constraints. Specifically, the system establishes indirect visual constraints throughout the entire process, only constructing direct photometric residuals in low-texture scenes to achieve local hybrid visual constraints, while initializing direct pixels through indirect poses. As a result, it can take full advantages of the visual constraints, enhancing the accuracy of localization. In addition, we design a novel mutual information-based direct photometric residual by introducing the illumination invariance property of mutual information, which improves the robustness of the direct method to illumination varying. As a consequence, the state estimator based on multi constraints is constructed, including pure indirect visual constraints or hybrid visual constraints, inertial residuals, and prior information. Comparative experiments on the public dataset and real-world flight experiments show that the proposed system outperforms state-of-the-art VIO frameworks in terms of both the accuracy and the robustness.