Understanding the Behavior of Data-Driven Inertial Odometry with Kinematics-Mimicking Deep Neural Network

In navigation, deep learning for inertial odometry (IO) has recently been investigated using data from a low-cost IMU only. The measurement of noise, bias, and some errors from which IO suffers is estimated with a deep neural network (DNN) to achieve more accurate pose estimation. While numerous studies on the subject highlighted the performances of their approach, the behavior of data-driven IO with DNN has not been clarified. Therefore, this paper presents a quantitative analysis of kinematics-mimicking DNN-based IO from various aspects. First, the new network architecture is designed to mimic the kinematics and ensure comprehensive analyses. Next, the hyper-parameters of neural networks that are highly correlated to IO are identified. Besides, their role in the performances is investigated. In the evaluation, the analyses were conducted with publicly-available IO datasets for vehicles and drones. The results are introduced to highlight the remaining problems in IO and are considered a guideline to promote further research.