Trainable Quaternion Extended Kalman Filter with Multi-Head Attention for Dead Reckoning in Autonomous Ground...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Extended Kalman Filter (EKF) is crucial for optimal control and mobile robot localization.
• Accurate noise modeling for EKFs is challenging due to dynamic environments and sensor uncertainty.
• Existing EKF methods struggle with unreliable default sensor noise values.

Purpose of the Study:

• To design a highly accurate, trainable EKF-based localization framework for autonomous ground vehicles (AGVs).
• To optimize EKF parameters using deep learning techniques for improved localization performance.
• To fuse IMU-based dead reckoning with LiDAR SLAM for enhanced AGV navigation.

Main Methods:

• Developed a trainable EKF framework utilizing inertial measurement units (IMUs) for AGV dead reckoning.
• Implemented Convolutional Neural Networks (CNNs), backward propagation, and gradient descent for parameter optimization.
• Designed a unique cost function to train the deep learning models and enhance EKF accuracy.
• Proposed a fusion strategy combining IMU-based localization with LiDAR SLAM estimation.

Main Results:

• Achieved highly accurate localization for autonomous ground vehicles (AGVs).
• Demonstrated improved EKF performance through trainable noise model optimization.
• Successfully fused IMU data with LiDAR SLAM for robust navigation.
• The framework showed general applicability to various IMU-aided robot localization tasks.

Conclusions:

• The proposed trainable EKF framework significantly enhances AGV localization accuracy.
• Deep learning optimization of noise models is effective for improving EKF performance in dynamic environments.
• Fusion of IMU and LiDAR SLAM provides a robust solution for autonomous navigation.
• The general design is adaptable for diverse robotic localization applications.