ResKACNNet: A Novel Backbone for Inertial Localization with Enhanced Long-Term Trajectory Modeling
Inertial localization is essential for positioning in environments where vision, GPS, or radar systems are not available. This paper introduces ResKACNNet, a new inertial localization network that overcomes the limitations of existing CNN-based methods in modeling long-term trajectory dependencies. By utilizing ChebyKAN as the backbone and incorporating an efficient kernel-based self-attention module (EKSA), the proposed method effectively captures complex trajectories and contextual information, leading to significant improvements in modeling long-term dependencies. Our experiments on five public datasets demonstrate a reduction in absolute trajectory error by 3.79%-42.32% compared to current baselines. Furthermore, we offer a preprocessed version of the TLIO dataset with gravity compensation, which has been experimentally proven to enhance inertial localization performance.
The dataset used in this research can be downloaded from the following link: https://pan.baidu.com/s/1CRPjyIHWFuqjU1rQQrcwCQ?pwd=mjbj Extraction Code: mjbj