Visual-Inertial-Wheel Odometry with Slip Compensation and Dynamic Feature Elimination.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Fusion

Background:

• Robot localization and odometry face challenges due to sensor uncertainties and wheel slippage.
• Visual-inertial-wheel odometry (VIWO) combines multiple sensor inputs for robust navigation.
• Existing VIWO systems struggle with accurate performance in challenging terrains and dynamic environments.

Purpose of the Study:

• To develop a novel data-driven approach for compensating wheel slippage in VIWO systems.
• To enhance the accuracy and robustness of robot localization and odometry.
• To improve the integration of visual and inertial measurements by addressing dynamic feature points.

Main Methods:

• Utilized Gaussian process regression (GPR) with deep kernel learning to model and mitigate slippage-induced errors.
• Incorporated long short-term memory (LSTM) layers for advanced error modeling.
• Developed a feature confidence estimator to handle dynamic feature points in visual data.
• Employed a multi-state constraint Kalman filter (MSCKF) for state estimation.

Main Results:

• The proposed method effectively compensates for wheel slippage, significantly improving localization accuracy.
• The system demonstrates enhanced robustness in challenging terrains and dynamic environments.
• Integration of GPR and LSTM layers led to superior performance compared to conventional VIWO systems.
• Experimental validation confirmed the effectiveness of the approach using real-world datasets.

Conclusions:

• The novel data-driven approach effectively addresses wheel slippage and dynamic feature point issues in VIWO.
• The enhanced VIWO system offers improved accuracy and robustness for autonomous navigation.
• This research contributes to the advancement of multi-sensor fusion and navigation technologies for unpredictable environments.