FOV Optimization

A fundamental assumption in robotic perception algorithms is that the sensor’s field-of-view (FoV) is fixed w.r.t the robot. This has led to research in areas such as active vision where robot planning is adapted to improve perception. More recently, there is an emerging class of sensors that can dynamically adjust their field of view during runtime, including PTZ cameras, sensors mounted on gimbals, RF radar systems, and others. This class of sensors decouple their orientation from that of the robot allowing for greater flexibility in algorithm design as well as separation of perception and planning al- gorithms. Such sensors require novel algorithms to calculate the sensor FoV for best perception. In this paper, we propose an online sensor view generation method over SO(3) that optimizes sensor’s rotation by maximizing the amount of visual features in the sensor’s limited FoV per Fisher Information metric. Since the amount of features can vary widely, our sensor view planning framework can be naturally extended to various perception applications such as localization, mapping, tracking and others.