Monocular Visual-Inertial Odometry with an Unbiased Linear System Model and Robust Feature Tracking Front-End

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Visual-inertial odometry (VIO) is maturing but faces accuracy-computation tradeoffs.
• Quaternion notation in VIO can cause confusion and hinder understanding.
• Existing VIO methods may struggle with feature matching outliers.

Purpose of the Study:

• To develop a VIO algorithm that optimizes both precision and computational efficiency.
• To address notation ambiguity in quaternion representations for VIO.
• To enhance the robustness of VIO systems against outliers.

Main Methods:

• A filter-based VIO solution using the multi-state constraint Kalman filter framework.
• Deduction of error state transition equations using Hamilton's quaternion notation for clarity.
• Development of a linear, closed-form formulation for easy implementation.
• Integration of a descriptor-assisted optical flow tracking front-end to handle outliers.
• Implementation of an automatic initialization procedure using static data.

Main Results:

• The proposed VIO algorithm achieves precision comparable to state-of-the-art methods.
• The new approach demonstrates significantly higher computational efficiency.
• The Hamilton quaternion notation simplifies understanding of the error state transition.
• The descriptor-assisted front-end effectively mitigates issues from feature matching outliers.

Conclusions:

• The developed VIO system offers a practical balance between accuracy and computational cost.
• The use of Hamilton quaternions and an improved front-end enhances VIO usability and robustness.
• This work contributes a more efficient and understandable VIO solution for researchers and practitioners.