9711528 Negahdaripour Vision-based automatic target tracking for recognition (ATR) is an important capability for intelligent or autonomous robotics systems. Applications include searching for (lost) objects, localization of nearby obstacles, mapping, docking with structures or moving objects surveillance, and surveying. Despite progress on the development of vision- based terrestrial and airborne ATR systems, realization of similar capabilities for submersible systems is a challenging problem due to the unique conditions of the undersea environment. This research effort involves a study of the vision-based ATR problem using a theory-based scientific approach, resting on physics-based modeling and active sensing paradigms. The goal is to provide an underwater vehicle system with the intelligence to perform the following functions autonomously: Explore immediate surroundings without collision with nearby obstacles. Locate objects of interest and fixate on them. Make intelligent moves and execute appropriate actions to efficiently examine the target and obtain useful information for recognition. Construct a knowledge base about the scene by integration of information from multiple views and different imaging conditions. Specifically, the researchers seek to develop techniques for obtaining information (e.g., three-dimensional shape, range, motion, reflectance properties) about objects in the scene from video data, according to: Physics-based models of underwater image formation, taking into account scene geometry and reflectance characteristics, illumination and viewing geometry, and ocean optical properties. Novel theories for adaptive online adjustment of sensor parameters (e.g., field of view, focus of attention), and positions and directions of viewing and scene illumination, to optimize data quality and information content. Mathema tical approach to registration and merging of visual information, obtained from various views and under different imaging conditions. ***
