Improving the accuracy of a low-cost inertial navigation system by adaptive Zero Velocity Updating using machine learning

Abstract In general, the Zero-Velocity updating method (ZUPT) using artificial neural networks has been widely used in pedestrian tracking detection systems using inertial sensors. On the other hand, in low-cost sensors such as MPU9250, the measurement bias is particularly large, so the motion state must be classified very much according to different speeds and amplitudes. We also need to have a pre-experience threshold that fits many of these states. Hence, we have proposed a rational new algorithm that, based on the work of the precursors, can provide high accuracy even in low-cost inertial sensors such as MPU9250. Here, we have developed a zero-velocity detector that uses a support vector machine to determine only the stationary state without classifying different motion states. The results of the experiment show that even in low-cost inertia sensors, a very accurate result is obtained with different speeds.