InertialNet: Inertial Measurement Learning for Simultaneous Localization and Mapping.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Simultaneous Localization and Mapping (SLAM) is vital for robot navigation.
• Determining 3D camera orientation in visual SLAM presents significant challenges.
• Existing methods often struggle with varying environmental conditions and visual complexities.

Purpose of the Study:

• To introduce InertialNet, an end-to-end network for predicting camera motion pose.
• To leverage the correlation between image sequences and Inertial Measurement Unit (IMU) signals.
• To develop a robust SLAM system independent of specific training set appearances.

Main Methods:

• Developed InertialNet, an end-to-end network integrating image and IMU data.
• Incorporated an optical flow substructure for environmental adaptability.
• Focused on inertial measurement learning for motion pose prediction.

Main Results:

• InertialNet demonstrated feasibility in predicting camera motion.
• The network showed stable predictions despite image blur, illumination changes, and low-texture scenes.
• Evaluations on EuRoC and custom datasets confirmed faster training convergence and fewer parameters.

Conclusions:

• InertialNet effectively predicts camera motion pose for enhanced SLAM.
• The network's design ensures robustness and adaptability to diverse environments.
• This approach offers an efficient solution for inertial measurement prediction in autonomous systems.