Non-iterative visual odometry

• Main Author: Yang, Zheng
• Other Authors: Xie Lihua
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: https://hdl.handle.net/10356/169487
• Institution: Nanyang Technological University

In 2017, a Non-Iterative Simultaneous Localization And Mapping system (NI-SLAM) was proposed by Nanyang Technological University. This system circumvents the instability issues of traditional feature-based methods in sparse and similar feature scenes by utilizing a novel frequency-domain pose estimation approach. Additionally, by employing Fast Fourier Transform (FFT) for frequency-domain correlation calculations and eliminating feature extraction, matching, and pose iteration optimization processes, NI-SLAM significantly improves the efficiency of pose estimation, resulting in a clear advantage in terms of speed compared to traditional methods. However, for achieving 6 degrees of freedom pose estimation, NI-SLAM relies on unconventional 9-axis Inertial Measurement Unit (IMU) for rotational correction before estimating the displacement. The 9-axis IMU utilizes magnetometers for directional estimation, which introduces errors in rotation estimation in scenarios affected by magnetic interference, such as being close to electric motors or wires where magnetometer readings may have significant inaccuracies. Furthermore, 9-axis IMUs are often expensive and require more power to operate compared to 6-axis IMUs, making them unsuitable for applications with strict cost and battery life requirements. Therefore, in this dissertation, we first propose a novel solution for rotation estimation instead of using a 9-axis IMU and combine it with a non-iterative displacement solver to develop a Visual Odometry (VO) solution called Depth Provided Rotation Estimation Non-Iterative Visual Odometry (DPRE NI-VO). In this approach, we reduce the hardware requirements of NI-SLAM to an RGB camera and a depth camera. By eliminating the constraints posed by the 9-axis IMU, our DPRE NI-VO becomes applicable to a broader range of indoor 6 degrees of freedom (DOF) pose estimation tasks. Secondly, we further extend the non-iterative pose estimation solution to ground texture-based localization tasks and propose Ground Texture-based Visual Odometry (GT NI-VO). In this approach, we remove the dependency on the depth camera, now requiring only a monocular camera to accomplish indoor and outdoor localization tasks. We provide detailed experiments for both proposed solutions in this dissertation and validate their accuracy and robustness on publicly available and self-collected datasets. Furthermore, we achieve superior robustness and accuracy performance compared to state-of-the-art algorithms through comparative analysis.