MagLand: Magnetic Landmarks for Road Vehicle Localization

Satellite-based systems are the most widespread solution for outdoor localization. However, they present well-known limitations in multipath environments and non-line-of-sight satellite conditions, e.g. tunnels, underground, urban canyons, and multilevel roads, being frequently combined with dead reckoning techniques. Inertial sensors present cumulative errors, and geomagnetic-field information is often distorted by strong local magnetic fields caused by road infrastructure. We turn this magnetic weakness into strength by proposing MagLand, an approach with detection and matching steps to leverage these anomalies as signatures for localization purposes. For anomaly detection, we adopt a window-based technique and apply supervised binary classification, choosing a random forest. We select one nearest centroid algorithm with dynamic time warping to match input data streams to reference signatures, providing guidelines to define and collect them. Real data experiments with off-the-shelf devices in challenging road scenarios show MagLand's feasibility with anomaly detection accuracy of 91% and matching with lane distinction of up to 93%. Magnetic landmarks can be extremely useful to address limitations of current localization systems and improve their performance, e.g. by providing an alternative in GPS limited areas, anchors for integrity monitoring, or resetting dead reckoning cumulative errors.