Robust Outlier-Adaptive Filtering for Vision-Aided Inertial Navigation.

Area of Science:

• Robotics and Autonomous Systems
• Navigation and Control
• Computer Vision

Background:

• Vision-aided inertial navigation systems (V-INS) are crucial for unmanned aerial vehicles (UAVs).
• Extended Kalman Filters (EKFs) are commonly used in V-INS but assume Gaussian noise, struggling with real-world non-Gaussian outliers.
• Outliers in sensor data can significantly degrade the reliability and robustness of V-INS.

Purpose of the Study:

• To develop an accurate and robust V-INS for UAVs that can handle unknown measurement outliers.
• To improve the performance of V-INS in realistic scenarios where data outliers are frequent.

Main Methods:

• Implemented a front-end feature correspondence method to reject initial vision outliers.
• Utilized a statistical test in the back-end filtering stage to detect remaining outliers.
• Employed variational approximation for Bayesian inference to compute optimal noise precision matrices for outlier adaptation.
• The integrated approach is termed 'outlier-adaptive filtering'.

Main Results:

• The proposed outlier-adaptive filtering framework significantly improves the accuracy of V-INS.
• The method demonstrates enhanced robustness in handling non-Gaussian measurement outliers.
• Flight dataset validation confirms the effectiveness of the developed V-INS framework.

Conclusions:

• Outlier-adaptive filtering is a critical advancement for robust V-INS in UAVs.
• The developed method addresses limitations of traditional EKF-based V-INS by effectively managing data outliers.
• This research provides a detailed treatment of outlier adaptation, enhancing navigation system reliability.