DL-Based Model Transformation Across Sonar Technologies for 3-D Target Reconstruction and Recognition
Abstract
Multi-beam forward-look (FL) sonar imaging has numerous applications in critical Navy operationswithin turbid waters, including butnot limited to search for, localization, and classificationof critical targets, and more generally scene characterization and modeling. To automate suchtasks, it is often instrumental to generate a 3-D model an object from a number of 2-D imagesfrom various viewpoints. To this end, the fundamental approach is to seek a 3-D object model thatcorresponds to synthesized 2-D sonar views that are most consistent with the data. Here, the keyassumption is the ability to generate such 2-D views, requiring a sonar simulator mimicking thecharacteristics of the specific s onar. Unfortunately, it is typically the case that some device-specificinformation is not publicly available and thus cannot be suitably modeled. To be specific, mostcommercial sonar systems are built according to different proprietary technologies, thus requiringsonar-specific image generators; a sonar manufacturer owning the proprietary technology seldomreleases detailed information, some of which play key roles in developing the device-dependentcomponents of the image generator. This implies that 1) a 3-D modeling method is effective to theextent that an available generator #sufficiently# mimics the sonar typeus ed to acquire the data;2) it may not be generalized for a new technology with different image formation characteristics,unless proper image generator adjustments are made. Additionally, the existing methods for 3-Dreconstruction from 2-D FL sonar views are ineffective for objects with complex shapes, e.g., withrandomly configured a rbitrary number o f p arts a nd ( or) s ignificant co ncavities. One ke y aspectof this effort rests on the premise that image formation models to establish device-specific generatorscan be effectively learned from any available model based on a different te chnology. In otherwords, the 2-D image variations of a target acquired by two different devices, due to sonar-specificinternal components, may not be readily modeled mathematically but can be learned from sufficientnumber o f diverse e xamples. This b enefits di rectly from powerful machine le arning (ML)techniques and the accelerated pace of new developments in deep learning (DL). Here, we makeuse of a mathematical DIDSON image formation model to explorethree DL-based techniquesfor model transformation to an Oculus sonar, to be further verified f or a BlueView s onar. Anotheraspect of the proposed project is to devise an effective method f or t he 3-D reconstructionof targets with high shape complexity from a multitude of 2-D views at known sonar poses. Here,we integrate an efficient hierarchical model-driven data acquisition procedure with a novel shapepruning scheme. Finally, we will investigate an efficient image matching method for model-basedrecognition through the generation of a rich image database of potential target classes producedby the Oculus image generator. The implementation and integration of our novel contributionsare expected to enhance AUV autonomy in carrying out the cited Navy operations, and the likes.At a total budget of $680,350, this three-year project will support 1) the PI (1.5 summermonths) for technical guidance, supervision and project management, two Ph.D. students (12-month stipend, annual health insurance, and first-year t uition), a nd a p art-time undergraduatestudent (420 hrs/year) for technical contributions, data collection efforts and experimental work;2) three annual trips to present the research findings at two professional meetings, and PI#s tripto ONR annual workshop for sponsored projects; 3) cost of resurfacing the water tank facilitythat will be utilized for extensive collection of data (first two years); 4) funds to purchase a sonarrotator for high-precision pose setting; and 5) supplies.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 08, 2024
- Source ID
- N000142412497
Entities
People
- Shahriar Negahdaripour
Organizations
- Office of Naval Research
- United States Navy
- University of Miami