Sonar image based advanced feature extraction and feature mapping algorithm for under-ice AUV navigation

Navigation and localisation of AUVs are challenging in underwater or under-ice environments due to the unavoidable growth of navigational drift in inertial navigation systems and Doppler velocity logs, especially in long-range under-ice missions where surfacing is not possible. Similarly, acoustic transponders are time consuming and difficult to deploy. Terrain Relative Navigation (TRN) and Simultaneous Localisation and Mapping (SLAM) based technologies are emerging as promising solutions as they require neither deploying sensors nor the calculation of distance travelled from a reference point in order to determine the location. These techniques require robust detection of the features present in sonar images and matching them with known images. The key challenge of under-ice image based localisation comes from the unstructured nature of the seabed terrain and lack of significant features. This issue has motivated the research project presented in this thesis. The research has developed technologies for the robust detection and matching of the available features in such environments. In aiming to address this issue, there are number of feature detectors and descriptors that have been proposed in the literature. These include Harris corner, Scale-Invariant Feature Transform (SIFT), Binary Robust Invariant Scalable Keypoints (BRISK), Speeded-Up Robust Features (SURF), Smallest Univalue Segment Assimilating Nucleus (SUSAN), Features from Accelerated Segment Test (FAST), Binary Robust Independent Elementary Features (BRIEF) and Fast Retina Keypoints (FREAK). While these methods work well in land and aerial complex environments, their application in under-ice environments have not been well explored. Therefore, this research has investigated the possibility of using these detector and descriptor algorithms in underwater environments. According to the test carried out with side-scan sonar images, the SURF and Harris algorithms were found to be better in repeatability while the FAST algorithm was found to be the fastest in feature detection. The Harris algorithm was the best for localisation accuracy. BRISK shows better immunity noise compared to BRIEF. SURF, BRISK and FREAK are the best in terms of robustness. These detector and descriptor algorithms are used for a wide range of varying substrates in underwater environments such as clutter, mud, sand, stones, lack of features and effects on the sonar images such as scaling, rotation and non-uniform intensity and backscatter with filtering effect. This thesis presents a comprehensive performance comparison of feature detection and matching of the SURF, Harris, FAST, BRISK and FREAK algorithms, with filtering effects. However, these detectors and descriptors have reduced efficiency in underwater environments lacking features. Therefore, this research further addressed this problem by developing new advanced algorithms named the 'SURF-Harris' algorithm, which combined Harris interest points with the SURF descriptor, and the 'SURF-Harris-SURF' algorithm which combined Harris and SURF interest points with the SURF descriptor, using the most significant factor of each detector and descriptor to give better performance, especially in feature mapping. The major conclusion drawn from this research is that the 'SURF-Harris-SURF' algorithm outperforms all the other methods in feature matching with filtering even in the presence of scaling and rotation differences in image intensity. The results of this research have proved that new algorithms perform well in comparison to the conventional feature detectors and descriptors such as SURF, Harris, FAST, BRISK and FREAK. Furthermore, SURF works well in all the disciplines with higher percentage matching even though it produces fewer keypoints, thus demonstrating its robustness among all the conventional detectors and descriptors. Even if there are a large number of features in a cluttered environment, it produces less matching compared to features that are distributed. Another conclusion to be derived from these results is that the feature detection and matching algorithms performed well in environments where features are clearly separated. Based on these findings, the comprehensive performance of combined feature detector and descriptor is discussed in the thesis. This is a novel contribution in underwater environments with sonar images. Moreover, this thesis outlines the importance of having a new advanced feature detector and mapping algorithm especially for sonar images to work in underwater or under-ice environments.