Comprehensive Perception in Marine Environments for Dynamic AnchoringOliver Ferdinand; Nick Rüssmeier; Andrej Lejman; Martin Günther; Friedemann Kammler; Frederic Theodor Stahl; Oliver Zielinski
In: Proceedings of OCEANS 2023. OCEANS MTS/IEEE Conference (OCEANS-2023), June 5-8, Limerick, Ireland, ISBN 979-8-3503-3226-1, IEEE, 2023.
For highly automatized or autonomous systems, where machines should cooperate and interact with humans on a high semantic level, establishing and maintaining hypothetical relationships between signals and symbols related to perceived objects is an important and challenging problem. Considering changes of physical and functional features of an object over time in dynamic and non-transparent environments, distinguishing objects is not just a matter of sensing. Here different objects may have a very similar appearance, yet they differ in their functionality and spatial context. For many of applications, e.g., a service robot to make sense of a mundane environment, it needs the ability to distinguish different objects and handle them appropriately, like individual tools in workshop environments, or vehicles like vessels in marine environments. Whenever a particular or individual object of a given set needs to be spatiotemporally recovered only from sensor data and prior knowledge about the object, an anchoring problem needs solving. Anchoring therefore enables stable perception of individual objects despite imperfect tracking possibilities, like during signal loss, noise, or coverage. The overall aim is the identification of typical and artificial intelligence relevant properties of moving objects in its typical environments by available and easily utilizable sensor techniques. The DFKI project CoPDA addresses the anchoring problem within comprehensive and ambitious settings, seeking to develop the main anchoring functions resulting in algorithms sets within the framework of Robot Operation System (ROS). The resulting software termed Dynamic Anchoring Agent (DAA) is suitable to derive, store and analyze essential information in real-time from broadband perception data. Within this paper, we shortly introduce the concept of the DAA and present an implemented realistic outdoor scenario in a harbor site. The dataset from an enhanced field sensor network, gathering field data of dynamic behaviors of individual sailing boats is described. It serves as basis for development, training, and evaluation of the DAA and is subsequently analyzed within object identification algorithms like YOLO or MaskR-CNN to show principal functionality for object perception and identification.