Estimating Dynamics On-the-Fly Using Monocular Video
We present an optimization framework to help estimate on-the-fly both the motion and physical parameters of an articulated multibody system using uncalibrated monocular image sequences. The algorithm takes video images of a physical system as input and estimates the motion together with the physical system parameters, given only the underlying articulated model topology. A valid initial pose of the system is found using a sequential optimization framework and used to bootstrap the successive pose estimation as well as estimation of physical system parameters (kinematic/geometric lengths as well as mass, inertia, damping coefficients). We also address the issue of robustly estimating a dynamically-equivalent system using partial state information (solely from noisy visual observations) and without explicit inertial parameter information. This framework results in a robust dynamically-equivalent system with good predictive capabilities when tested on a double pendulum system.