Extended Kalman filter-based methods for pose estimation using visual, inertial and magnetic sensors: comparative...

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Accurate ego-motion estimation is crucial for mobile robots and autonomous systems.
• Integrating visual and inertial/magnetic data offers potential for improved pose tracking.
• Existing methods often face challenges with sensor drift and computational complexity.

Purpose of the Study:

• To develop and compare two Extended Kalman filters (EKFs) for fusing monocular vision and Inertial Measurement Unit (IMU) data.
• To evaluate the performance of a Direct Linear Transformation (DLT)-based EKF against an error-driven EKF and a purely IMU-based EKF for ego-motion estimation.
• To assess accuracy, robustness, and computational complexity in various experimental conditions.

Main Methods:

• Developed two EKFs: one leveraging Direct Linear Transformation (DLT) for visual ego-motion, and another using projection errors (error-driven).
• Both EKFs fused measurements from a monocular vision system and an IMU rigidly attached to the camera.
• Compared the fused-sensor EKFs against a purely IMU-based EKF for orientation estimation.

Main Results:

• The DLT-based EKF achieved superior accuracy, with orientation root mean square errors (RMSEs) of 1° and position RMSEs of 3.5 mm.
• The error-driven EKF yielded orientation RMSEs of 1.5° and position RMSEs of 10 mm.
• A purely IMU-based EKF resulted in orientation RMSEs of 1.6°, demonstrating the benefit of sensor fusion.

Conclusions:

• The DLT-based EKF provides highly accurate ego-motion estimation by effectively fusing visual and IMU data.
• While less robust to visual feature loss, the DLT-based EKF offers comparable computational complexity to the error-driven approach.
• Sensor fusion significantly enhances pose estimation accuracy compared to using an IMU alone.