MKC-LSSVM: A Robust Motion Classification for ZUPT of Foot-mounted Inertial Navigation

The zero-velocity update (ZUPT) technique is crucial for mitigating the drift in foot-mounted inertial navigation systems (FMINS). However, the accuracy of zero-velocity detection (ZVD) is highly sensitive to motion types, which limits its effectiveness in dynamic scenarios. To address this challenge, we propose a novel motion classifier called the least-squares support vector machine based on multi-kernel correntropy (MKC-LSSVM), which adjusts the detection threshold of the widely used stance hypothesis optimal estimation (SHOE) detector according to motion types. Unlike traditional LSSVM, which uses a quadratic loss function, our approach employs a multi-kernel-correntropy-based loss function, enhancing robustness against sensor noise and outliers. The new non-convex optimization problem is transformed into two convex subproblems using the half-quadratic optimization method and solved via an alternating iterative algorithm. Furthermore, the whale optimization algorithm (WOA) is utilized to optimize the free parameters of MKC-LSSVM, ensuring optimal performance. Experimental results on public mixed-motion datasets demonstrate that the proposed MKC-LSSVM achieves superior motion classification accuracy compared to existing methods. By incorporating MKC-LSSVM, the new FMINS framework significantly enhances positioning accuracy through improved ZUPT capability. This work not only advances the motion classification for FMINS but also provides a robust classifier for complex dynamic applications.