SensorSLAM: An investigation into sensor parameter estimation for SLAM

Mapping technology is an essential component of autonomous robotic systems. The ability of a vehicle to establish its location with respect to some environmental representation allows the vehicle to navigate and reason about the environment. This process of estimating vehicle motion while fusing sensor data to acquire a map of the environment is called simultaneous localization and mapping or SLAM. Existing SLAM formulations are unable to effectively address the problem of non-stationary sensor noise in the sensor model. Yet orientation or location dependent noise sources are commonplace in real-world scenarios. The use of standard noise models (which assume stationary noise) leads to instability or reduced accuracy of the resulting map within traditional SLAM solutions. By parameterizing the non-stationary aspects of the noise model and estimating these parameters simultaneously with the map and sensor location, a stochastic formulation for SLAM is developed.;The general Bayesian framework developed in this dissertation is applicable in many domains, however this work focuses on the underwater environment. Stereo video cameras and inertial sensors are utilized to solve SLAM in the underwater environment. The resulting algorithm is used to recover models of complex underwater structures. Incorporating a non-stationary noise model within a Bayesian SLAM formulation reduces the error in the resulting trajectory and map by simultaneously estimating the location-based (non-stationary) noise in the environment, the trajectory of the sensor, and an accurate map of the environment.