Probabilistic Object Tracking Using Quantified Camera Uncertainty Parameters in a Binocular System

As the Navy becomes increasingly autonomous and attempts to lower its radio frequency (RF) signature, the need for improved optically-driven systems increases. Research on a binocular camera system's ability to track an object in three-dimensional (3D) space using statistical uncertainty supports this effort. By leveraging ground truth data provided by USNA's Vision Integration In Polymanual and Experimental Robotics (VIPER) lab, the binocular system's uncertainty parameters are experimentally estimated, and the overall uncertainty of the system's ability to track an object is approximated. Unlike traditional tracking methods, this process produces a probability “cloud” around the the object of interest, providing confidence intervals around the target of interest. This research supports sensor-based autonomy in two ways: (1) enabling tracking systems to use probabilistic information to establish confidence regions for risk assessment and decision making;; (2) provide tracking uncertainty in automated image labeling to improve the training efficacy of neural networks.