Pedestrian Localization on Topological Maps with Neural Machine Translation Network

Pedestrians and vehicles tend to follow certain paths even in the case of no road or walkable path is prescribed. This fact makes topological representation of maps very important in navigation and localization applications. When infrastructure-based systems (e.g., GPS, Wi-Fi) are available, the problem is referred to as map-matching where the goal is to improve localization accuracy. On the other hand, localization becomes a problem in the lack of initialization for infrastructure-free systems (e.g., dead reckoning via odometry). For example, a pedestrian inertial navigation system (INS) can be used to generate drift-prone trajectories that would perfectly turn into graphs (consisting of nodes and edges) after heuristic corrections such as zero-velocity update (ZUPT), heuristic drift reduction (HDR) and motion direction aid by Earth's magnetic field. Most of the approaches use probabilistic generative methods to solve sub-graph traversal problem. This research adopts a neural network structure that is originally designed for language translation purposes and modifies it to achieve pedestrian localization on topological maps. Different feature combinations (angle vs. edge length) and number of traversed nodes are tested in the experiments. A minimum accuracy of 95% is achieved with all feature configurations while traversing only six nodes on the map.