Contributions to robotic underwater perception and mapping based navigation

Abstract

In recent years, one of the fundamental challenges of the research on robotics is to obtain robust and efficient mechanisms for modelling increasingly complex environments, using mobile robots for their exploration. Nowadays, underwater vehicles are increasingly being used in complicated and rigid environments like oceans, harbors or at dams, such as using the on-board Side Scan Sonars (SSSs) of the Autonomous Underwater Vehicles (AUVs) to image the seabed. And underwater robotic mapping has been an active research area in robotic research community. Localization and mapping are the fundamental abilities for underwater robots to carry out exploration and searching tasks autonomously. Autonomous localization and mapping requires a vehicle to start at an unknown location in an unknown environment and then to incrementally build a map of landmarks present in this area while simultaneously using this map to compute absolute vehicle position. The problem of solving both map building and robot localization is addressed by Simultaneous Localization and Mapping (SLAM). The main focus of this thesis is on extracting features with Sound Navigation And Ranging (SONAR) sensing for further robotic underwater landmark-based SLAM navigation. First and foremost, a detailed overview of currently used and popular solutions for the underwater SLAM problem is presented and characteristics like accuracy, robustness, computational complexity, etc. are compared. Besides, different types of commonly used map representations are compared regarding their suitability for a priori map localization, in particular with regard to large-scale underwater SLAM based navigation, which are computational requirements, reliability, robustness, etc. In our case, consider the sparse spatial distribution of the marine features, thus the landmark map is chosen to represent the underwater region to be explored. According to the characteristics of sonar images, we propose an improved Otsu Threshold Segmentation Method (TSM) for fast and accurate detecting underwater objects of various feature shapes, including a shipwreck, a branch and a plastic mannequin in this thesis. For all the SSS and Forward Looking Sonar (FLS) images of different resolutions and qualities presented in this thesis, simulation results show that the computational time of our improved Otsu TSM is much lower than that of the maximum entropy TSM, which achieves the highest segmentation precision than other three classic TSMs including the traditional Otsu method, the local TSM and the iterative TSM. Experimental results justify that the improved Otsu TSM maintain more complete information and details of the objects of interests after segmentation, also the effect of noise, the holes and gaps in the foreground objects are greatly reduced. Furthermore, the MIxed exponential Regression Analysis (MIRA) method is presented for handling the echo decay in sonar images. Our MIRA model is compared with the Dark Channel Prior (DCP) model, which is an adaption to well-known fog removal technique to sonar imaging, in terms of one ping, local similarity and global quality analysis between a landmark and its 〖180〗^°- rotated counterpart. Simulation results prove that the proposed MIRA approach has superior normalization performances. In addition, a detailed state of the art underwater feature matching methods are summarized and compared, including the classic Harris, Speeded Up Robust Features (SURF), Binary Robust Invariant Scalable Keypoints (BRISK), Scale Invariant Feature Transform (ASIFT) feature detector. Consider their suitability for subsea sonar map fusion, in terms of computational requirements, reliability, accuracy, etc. We propose to employ the underwater wireless sensor network (UWSN) into the sonar map fusion task, and present UWSN-based Delaunay Triangulation (UWSN-DT) algorithm for enhancing the performances of sonar map fusion accuracy with low computational costs. Experimental results justify that the UWSN-DT approach works efficiently and robustly, especially for the subsea environments where distinguishable feature points are few and difficult to be detected. Most importantly, as a result of the segmentations, the centroids of the main extracted regions are computed to represent point landmarks which can be used for navigation, e.g., with the help of our newly proposed Augmented Extended Kalman Filter (AEKF)-based robotic underwater SLAM algorithm, which stores the robot pose and the map landmarks in a single system state vector, and estimates the state parameters by using a iterative, estimation-update process, which besides a prediction, an update stage (as well as in the classic Extended Kalman Filter), includes a newly proposed augmentation stage. Several MATLAB simulated experiments are carried out for both our proposed AEKF-SLAM based navigation algorithm and classic FastSLAM 2.0 for dense loop mapping and line mapping, where environmental landmarks include not only the calculated centroids of the shipwreck, branch and plastic mannequin, but also those centroids of certain parts of the background detected by the proposed improved Otsu TSM in Section 3. Simulation results prove that for both dense loop mapping and line mapping experiments, our AEKF-SLAM based robotic navigation approach has the best performances of localization and mapping accuracy with relatively low computational costs. The AEKF achieves more precise and robust estimations of the robot pose and the landmark positions on account of the landmark positions detected by our improved Otsu TSM, than those derived by the maximum entropy TSM. In all, our presented AEKF-based robotic underwater SLAM algorithm achieves reliable detection of cycles in the map and consistent map update on loop closure.