Under-ice relative pose estimation and ice anomaly mapping with an unmanned underwater vehicle
MetadataShow full item record
Under-ice regions in both the Arctic and Antarctica are of great interest to many domains of science including biology, climate science, and planetary science. Great environmental and technical challenges face researchers when attempting to gather data from the polar under-ice regions of Earth for these applications. The harsh environments encountered both above and below the polar ice limit the use of human divers and manned submersibles in such data collection efforts. However, recent technological advances have provided the means for data collection using unmanned underwater vehicles (UUVs) beneath the ice. New challenges are encountered using this unmanned technology including deployment, recovery, risk mitigation, navigation, and mapping. Presented in this dissertation are methods developed to help aid in navigation and mapping tasks in these under-ice environments. Specifically, development of computer vision methods using acoustic and optical imaging sensors (especially through sensor fusion) is used to help aid in under-ice UUV motion estimation. In addition to algorithms which can aid in the navigation problem, additional computer vision methods for ice texture and ice anomaly mapping are also developed and presented herein. The methods presented here utilize low-cost sensors already onboard many UUV platforms, and do not require expensive external infrastructure or setup effort. First, a relative pose estimation method, used to help track a vehicle's position over a trajectory, is presented using a novel combination of optical flow-based computer vision methods with forward-looking sonar data, assuming a rigid motion model between frames. The use of computer vision techniques with sonar data is uncommon due to the high noise levels from this sensor, as well as the changing appearance of objects between frames. In addition to the use of a sonar sensor to aid in vehicle motion estimation, the use of camera sensors, also commonly onboard UUVs, is presented to provide additional independent relative vehicle pose estimates. While monocular camera relative pose methods are commonly applied in feature-rich environments such as terrestrial urban and underwater coral reef locations, such methods are not common in featureless environments such as that found under the ice. In order to overcome these challenges, a novel adaptation of this approach is presented here using contrast enhancement and low feature detection thresholds, along with robust feature matching to eliminate outliers during motion model estimation. While relative pose estimation using either a sonar or camera sensor provides valuable information for autonomous navigation in a UUV, fusion of these two complementary sensors can result in a much stronger and more robust trajectory estimate. A novel sonar and camera sensor fusion approach is presented here using a factor graph framework to combine estimates from these noisy but partially redundant sensors. In this case, the camera estimates provide additional degrees of motion not possible using a sonar sensor, but the translational camera estimates contain a scale factor ambiguity inherent to the camera sensor. While the sonar estimates are limited to only three degrees of motion, no scale factor ambiguity is encountered, and absolute translational motion can be estimated. Fusion of these two sensors can be used to leverage the strengths of each sensor to overcome the individual weaknesses and provide much more robust overall vehicle motion estimates. Finally, it is not only useful to provide an estimate of the location of a UUV during data collection, but also to automatically estimate ice texture and to flag frames with possible anomalies of interest present against the ice background. This can help aid human vehicle operators and scientists in data collection and post analysis of these large, mostly featureless, datasets. A method for estimating ice texture using point features is presented here, along with point feature- and hue-based methods for anomaly detection and mapping. Methods such as those developed in this dissertation can help aid vehicle operators and scientists in the difficult tasks of navigation and mapping for under-ice exploration, and can eventually provide an autonomous means for such data collection. The algorithms developed here can further the mission capabilities of current under-ice vehicle platforms to enable further exploration of these remote areas of the planet.