Real-time camera tracking using a particle filter combined with unscented Kalman filters

摘要

Real-time camera tracking is steadily gaining importance due to the drive from various applications, such as augmented reality, three-dimensional structure estimation/modeling, and mobile computing environment. However, tracking a monocular camera in an unknown environment is not a trivial work. We describe a real-time camera tracking framework designed to track a monocular camera in a workspace. In particular, we focus on integration of a bundle of nonlinear filters to achieve robust camera tracking and scalable feature mapping, which can extend to larger environment. The basic idea of the proposed framework is that a particle filter-based camera tracking is connected to independent feature tracking filters, which have fixed-state dimension. In addition, every estimate required for template prediction is obtained from the independent feature estimators so that the template prediction can be maintained without additional framework for the template state estimation. We split the camera tracking and feature mapping into two separate tasks, and they are handled in two parallel processes. We demonstrate the effectiveness of the proposed approach within a desktop environment in real time.