Avnet: learning attitude and velocity for vehicular dead reckoning using smartphone by adapting an invariant EKF

Abstract Smartphone-based vehicle navigation has become the primary choice in everyday life due to low cost and real-time traffic updates. However, for vehicle navigation applications in the Global Navigation Satellite System (GNSS)-denied scenario such as parking lot and tunnel, it is quite difficult to maintain robust and continuous positioning based on consumer-grade sensors. In this paper, a novel method is proposed for accurate vehicular dead-reckoning based only on a smartphone inertial measurement unit. Robust vehicle dead reckoning can improve positioning performance in GNSS-degraded areas or where high-precision positioning sources are available at low-frequencies. The key components of the method are a Kalman filter with data-driven parameters adapter and a deep neural network that provides data-driven measurement estimation. A combined convolutional neural network and gated recurrent unit deep learning network, termed AVNet, is proposed to estimate the attitude and velocity of the vehicle. The learned measurements are integrated into an invariant Kalman filter to estimate Three-Dimensional (3D) attitude, velocity and position. The method was tested on custom datasets collected in a parking lot, and a 0.4 % relative horizontal translation error was achieved on average.