On environmental adaptation in GNSS-based integrated navigation systems
University of New Brunswick
The most efficient positioning and navigation solutions for mass-market applications are based on integration of data collected with several sensors installed on a user platform. The ability of the system, both to measure a variety of physical effects and to generate infrastructure-induced measurements, allows us to exploit strengths and to compensate, to an extent, for vulnerabilities of individual sensors. GNSS remains a key building block of integrated navigation systems; however, in urban scenarios it is negatively affected by reflection, attenuation, and blockage of radio signals. If unaccounted for, the GNSS performance variation caused by the above effects may result in sub-optimal state estimation. In a pursuit of a design allowing for a GNSS-based integrated navigation system to automatically adjust its parameters with respect to the surrounding GNSS environment, the following aspects were investigated. First, inaccurate specification of noise covariances in a Kalman filter may lead to a solution degradation. The novel concept of GNSS environment mapping has been developed, to allow a state estimation filter to adjust its measurement noise by relying on crowd-sourced GNSS measurement statistical representation over an urban area of operation, instead of relying solely on data available on an individual platform. Application of the concept leads to increased coordinate determination accuracy and to a faster solution re-convergence. By training a random forest model, the GNSS environment map availability is extended to areas for which no crowd-sourced GNSS data is available. Second, to decrease vulnerability of a minimum error variance estimator to inaccurate stochastic modelling, a filter could be extended with a worst-case error minimization criterion, such that it makes no assumptions on noise properties. In the new solution to continuously maintain balance between the two estimation criteria, a reinforcement learning framework has been introduced into the GNSS-based integrated navigation engine. Practical results show the capability of such a model to progressively self-improve while encountering diverse GNSS signal propagation scenarios. These novel developments have been tested with several configurations of GNSS-based integrated navigation engines; a 13% and 17% absolute accuracy improvement in the tightly- and loosely-coupled integration modes respectively is demonstrated. The identified shortcomings of the proposed techniques and the recommendations for further developments are provided.