LKAN: A Kolmogorov-Arnold Network-Based Framework with Long-History Statistical Regularization for IMU Trajectory...

Area of Science:

• Robotics and Autonomous Systems
• Sensor Fusion
• Machine Learning for Navigation

Background:

• Inertial Measurement Units (IMUs) are crucial for indoor trajectory estimation.
• Cumulative drift and nonlinear dynamics pose significant challenges for IMU-based localization.
• Existing methods struggle with high-precision, real-time indoor positioning.

Purpose of the Study:

• To introduce LKAN, an end-to-end framework for high-precision indoor trajectory estimation using IMUs.
• To enhance robustness and mitigate error accumulation in IMU-only localization.
• To leverage advanced neural network architectures for improved performance.

Main Methods:

• Integration of Kolmogorov-Arnold Network (KAN) with Long-History Statistical Regularization (LHSR).
• Development of a KANmer encoder fusing Multi-Head Self-Attention with KAN for temporal dependencies and nonlinear features.
• Implementation of a training-only LHSR mechanism to enforce historical statistical consistency and suppress feature drift.

Main Results:

• LKAN significantly outperforms state-of-the-art methods in IMU-only pedestrian localization.
• Achieved an Absolute Trajectory Error (ATE) of 2.04 m and Relative Trajectory Error (RTE) of 2.72 m on the iIMU-TD dataset.
• Demonstrated substantial error reduction (33.8% ATE, 31.1% RTE) compared to the second-best method, ResT-IMU.

Conclusions:

• LKAN provides a reliable, high-precision solution for real-time IMU-based positioning in complex indoor environments.
• The proposed framework effectively mitigates error accumulation inherent in IMU data.
• LKAN represents a significant advancement in IMU-only indoor localization technology.