NIN-VINS: Neural Inertial Navigation Aided Visual-Inertial System for Pedestrian Dead Reckoning

Traditional visual-inertial system track 6DoF motion based on IMU kinematic model and visual feature measurements. However, consumer-grade IMU faces the challenge of large cumulative error in the integration process resulted by sensor bias and noise. Visual navigation also relies on visual tracking that is constant and reliable, it can be difficult to do when there is a lack of or confusing visual information. Because of the complexity of human mobility, using a visual-inertial system to estimate pedestrian stance has presented new obstacles. We propose a visual-inertial pedestrian pose estimation system with data-driven inertial navigation assisted in this research. The neural inertial navigation system makes better use of IMU data to learn the potential motion mode of the human body, lessen the visual-inertial system’s reliance on visual data, and deliver more accurate pedestrian pose estimation results. Meanwhile, we make the neural inertial navigation model compatible with NVIDIA TensorRT runtime in order to improve its efficiency. The experimental results based on multiple open-source dataset and our self-collected data show that NIN-VINS provide higher accuracy compared with traditional visual-inertial system.