Theory, Design, and Implementation of Landmark Promotion Cooperative Simultaneous Localization and Mapping

Abstract

Simultaneous Localization and Mapping (SLAM) is a challenging problem in practice, the use of multiple robots and inexpensive sensors poses even more demands on the designer. Cooperative SLAM poses specific challenges in the areas of computational efficiency, software/network performance, and robustness to errors. New methods in image processing, recursive filtering, and SLAM have been developed to implement practical algorithms for cooperative SLAM on a set of inexpensive robots.

The Consolidated Unscented Mixed Recursive Filter (CUMRF) is designed to handle non-linear systems with non-Gaussian noise. This is accomplished using the Unscented Transform combined with Gaussian Mixture Models. The Robust Kalman Filter is an extension of the Kalman Filter algorithm that improves the ability to remove erroneous observations using Principal Component Analysis (PCA) and the X84 outlier rejection rule. Forgetful SLAM is a local SLAM technique that runs in nearly constant time relative to the number of visible landmarks and improves poor performing sensors through sensor fusion and outlier rejection. Forgetful SLAM correlates all measured observations, but stops the state from growing over time. Hierarchical Active Ripple SLAM (HAR-SLAM) is a new SLAM architecture that breaks the traditional state space of SLAM into a chain of smaller state spaces, allowing multiple robots, multiple sensors, and multiple updates to occur in linear time with linear storage with respect to the number of robots, landmarks, and robots poses. This dissertation presents explicit methods for closing-the-loop, joining multiple robots, and active updates. Landmark Promotion SLAM is a hierarchy of new SLAM methods, using the Robust Kalman Filter, Forgetful SLAM, and HAR-SLAM.

Practical aspects of SLAM are a focus of this dissertation. LK-SURF is a new image processing technique that combines Lucas-Kanade feature tracking with Speeded-Up Robust Features to perform spatial and temporal tracking. Typical stereo correspondence techniques fail at providing descriptors for features, or fail at temporal tracking. Several calibration and modeling techniques are also covered, including calibrating stereo cameras, aligning stereo cameras to an inertial system, and making neural net system models. These methods are important to improve the quality of the data and images acquired for the SLAM process.