Stereo Direct Sparse Visual-Inertial Odometry with Efficient Second-Order Minimization.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Visual-inertial odometry (VIO) is crucial for autonomous systems but struggles with initialization, illumination changes, and sensor fusion.
• Existing VIO methods face limitations in accuracy and robustness due to these challenges.

Purpose of the Study:

• To propose a novel stereo direct sparse visual-inertial odometry system.
• To overcome key challenges in VIO, including initialization sensitivity, dynamic illumination, and multi-sensor fusion.

Main Methods:

• The system utilizes a direct method approach with a depth initialization module based on visual-inertial alignment.
• It employs efficient second-order minimization (ESM) to jointly optimize camera poses and sparse scene geometry by minimizing photometric and inertial errors.
• Measurement preintegration and a marginalization module are used for IMU data processing and computational efficiency.

Main Results:

• The proposed system demonstrates state-of-the-art performance on the KITTI and EuRoC datasets.
• Experimental results confirm high accuracy and robustness in challenging visual odometry scenarios.

Conclusions:

• The developed stereo direct sparse visual-inertial odometry system effectively addresses critical VIO challenges.
• The approach offers a robust and accurate solution for autonomous navigation systems.