| Summary: | Work in the polar regions of our planet is unavoidably linked with hazards such as drift ice. Increased presence, fueled by economic interests in the Arctic, has for several decades called for research in the field of ice management. The field deals with the detection, tracking and forecasting of ice, but also the physical actions taken to avoid collisions. Managing ice is of great importance to polar ventures but predicting ice movement has proven difficult, concluding that observations are essential for tracking. Ice management is applicable not only for stationary installations, but has also been studied for the protection of ship routes in the Arctic. A moving object may be not be able to rely solely on satellite imagery, as these generally have limited availability and coverage, with sample times in the order of days. Recent studies have introduced the use of Unmanned Aerial Systems (uas's) as a supplement to local sensors such as ship- or ground-based radar. This thesis aims to provide a detailed insight into the design of a scalable sea ice tracking system with components ranging, from data collection, classification and tracking, to the feedback of previous tracking results into the formation of new paths for multiple uas agents to collect new observations. Beyond applying the uas as a sensor platform, examples are also given in this thesis of the use of machine learning algorithms and background modeling to extract and classify sea ice detections from Synthetic Aperture Radar (sar) and ground-based radar. In the business of tracking ice objects in the Arctic, objects are tracked over a geographically vast area. While each observation covers only a relatively limited area, a complete system needs to handle large number of ice objects. Commonly, Multiple Target Tracking (mtt) algorithms scale poorly with target numbers, which poses a problem to the large scale of our scenario. This thesis goes into detail on selected approaches for enabling large-scale mtt. First, we introduce spatial indexing for a fast partitioning scheme for the Multiple Hypothesis Tracker (mht) and Labeled Multi- Bernoulli (lmb) algorithms. Further, we propose a novel formulation of the lmb filter designed to simplify its implementation. The resulting filters are detailed, implemented and applied to sea ice tracking scenarios. To utilize the information available from the tracker, we explore two applications in particular. First, we study probabilistic modeling of current and wind velocities based on tracker data and Gaussian fields. Second, we propose the use of the Probability Hypothesis Density (phd) for informed planning of uas flight paths. The phd—the density of expected number of tracked objects—can be efficiently extracted from the same data structures used in e.g. the lmb filter, and is thus used to form a common “language” between the two algorithm families of target trackers and path planners. A proof of concept multi-agent path planner is developed and published open-source along with implementations of the mht and lmb filters. digital fulltext not avialable
|
|---|