Learning Gaussisan noise models from high-dimensional sensor data with deep neural networks

Abstract

While measurement covariances are often taken to be constant in many robotic state estimation systems, many sensors exhibit different interactions with their environment. Accurate covariance estimation allows graph-based estimation techniques to better optimize state estimates by reasoning about the utility of different methods relative to each other. This thesis describes a method of learning compact feature representations for real-time covariance estimation. A direct log-likelihood optimization technique is used to train a deep convolutional neural network to predict the covariance matrix of a Gaussian measurement model, given representative data. This method is algorithm-agnostic, and therefore does not require the handcoding of representative features. Quantative results are presented, showing that improved measurement covariances on a frame-to-frame visual odometry system reduce trajectory errors after a loop closure is applied.