Five-dimensional local positioning using neural networks

Abstract

In this thesis, a method for real-time transmitter localization is evaluated. An existing system has acted as testbed for the evaluation. This system uses an electromagnetic transmitter and a receiver board with 16 antennas. The antenna values are used to recover the transmitters position and two angles, the five dimensions. The proposed solution is an inverse modelling feed-forward neural network, a multilayer perceptron, which is trained and evaluated with the use of the TensorFlow library. The project resulted in a purely software based estimator which requires no change to the testbed and can act as a drop in replacement for the previous algorithm. The new estimator has accomplished improvements in estimation speed (more than 100× faster), expansion of the volume in which the position can be recovered (27× larger), enlarged range of angles (10% per axis) and has improved the precision of the position estimates (error at the 95th percentile reduced to ~ 1/3 of the previous implementation). The new algorithm is a substantial improvement on the previous implementation, enabling new use cases for the system.