Maximum Likelihood Fusion Models

Abstract

In the absence of a global frame of reference, the ability to fuse data collected by multiple mobile agents that operate in separate coordinate systems is critical for enabling autonomy in multi-agent navigation and perception systems. Of particular interest is the ability to fuse rigid body metric environment models in order to construct a global model from the data collected by each agent. This thesis presents a data fusion approach for combining Gaussian metric models of an environment constructed by multiple agents that operate outside of a global reference frame. Common landmarks are combined using a nonlinear least squares approximation, which yields an exact solution under the assumption of isotropic covariance. Rigid body transform parameters and common landmarks are found using a hypergraph registration approach. The approach demonstrates a robustness to outliers in registration by incorporating unit quaternions to reject outliers on a unit sphere. The performance of the approach is evaluated using experimental benchmark datasets collected in natural and semi-structured environments with camera and laser sensors.