L5IN^+: From an Analytical Platform to Optimization of Deep Inertial Odometry

Fifth generation of mobile communications (5G) and Deep Neural Networks (DNN) are two important technologies, which will enable new functions in the field of indoor positioning. This could be seen as the second major development after the innovation of smartphones, as a GNSS/INS alternative for indoor, location based applications. Optimization methods which work as a corrector, and as the uncertainty assessment for real life applications, guided us through the next level of challenges. In this paper, we have opened a novel interpretation of a deep network for inertial odometry which is robust to noisy labelled data that was detected from a 5G network. We also designed and developed analytical platform, which is considered a data collector and cellular positioning simulation. This platform was used to provide the input for the learning and optimization algorithms. The simulation website is implemented and available online under simulation2evaluation.herokuapp.com for researchers to generate ground truth trajectories and simulated cellular measurements with assigned quality and exact error values. We have proposed two approaches: (1) deep inertial odometry based on predicting velocity vector elements or relative positions and (2) Kalman Filtering to use, combine and test the absolute positions with the relative ones from the first approach. We finally provide numerical results of our experiments and a discussion of the effectiveness of our approaches.