Methods, Models, and Datasets for Visual Servoing and Vehicle Localisation

Abstract

Machine autonomy has become a vibrant part of industrial and commercial aspirations. A growing demand exists for dexterous and intelligent machines that can work in unstructured environments without any human assistance. An autonomously operating machine should sense its surroundings, classify different kinds of observed objects, and interpret sensory information to perform necessary operations.

This thesis summarizes original methods aimed at enhancing machine’s autonomous operation capability. These methods and the corresponding results are grouped into two main categories. The first category consists of research works that focus on improving visual servoing systems for robotic manipulators to accurately position workpieces. We start our investigation with the hand-eye calibration problem that focuses on calibrating visual sensors with a robotic manipulator. We thoroughly investigate the problem from various perspectives and provide alternative formulations of the problem and error objectives. The experimental results demonstrate that the proposed methods are robust and yield accurate solutions when tested on real and simulated data. The work package is bundled as a toolkit and available online for public use. In an extension, we proposed a constrained multiview pose estimation approach for robotic manipulators. The approach exploits the available geometric constraints on the robotic system and infuses them directly into the pose estimation method. The empirical results demonstrate higher accuracy and significantly higher precision compared to other studies.

In the second part of this research, we tackle problems pertaining to the field of autonomous vehicles and its related applications. First, we introduce a pose estimation and mapping scheme to extend the application of visual Simultaneous Localization and Mapping to unstructured dynamic environments. We identify, extract, and discard dynamic entities from the pose estimation step. Moreover, we track the dynamic entities and actively update the map based on changes in the environment. Upon observing the limitations of the existing datasets during our earlier work, we introduce FinnForest, a novel dataset for testing and validating the performance of visual odometry and Simultaneous Localization and Mapping methods in an un-structured environment. We explored an environment with a forest landscape and recorded data with multiple stereo cameras, an IMU, and a GNSS receiver. The dataset offers unique challenges owing to the nature of the environment, variety of trajectories, and changes in season, weather, and daylight conditions. Building upon the future works proposed in FinnForest Dataset, we introduce a novel scheme that can localize an observer with extreme perspective changes. More specifically, we tailor the problem for autonomous vehicles such that they can recognize a previously visited place irrespective of the direction it previously traveled the route. To the best of our knowledge, this is the first study that accomplishes bi-directional loop closure on monocular images with a nominal field of view. To solve the localisation problem, we segregate the place identification from the pose regression by using deep learning in two steps. We demonstrate that bi-directional loop closure on monocular images is indeed possible when the problem is posed correctly, and the training data is adequately leveraged.

All methodological contributions of this thesis are accompanied by extensive empirical analysis and discussions demonstrating the need, novelty, and improvement in performance over existing methods for pose estimation, odometry, mapping, and place recognition.