Visual motion estimation and tracking of rigid bodies by physical simulation

摘要

This thesis applies knowledge of the physical dynamics of objects to estimating object motion from vision when estimation from vision alone fails. It differentiates itself from existing physics-based vision by building in robustness to situations where existing visual estimation tends to fail: fast motion, blur, glare, distractors, and partial or full occlusion. A real-time physics simulator is incorporated into a stochastic framework by adding several different models of how noise is injected into the dynamics. Several different algorithms are proposed and experimentally validated on two problems:
\n
\n motion estimation and object tracking.
\n
\nThe performance of visual motion estimation from colour histograms of a ball moving in two dimensions is improved considerably when a physics simulator is integrated into a
\nMAP procedure involving non-linear optimisation and RANSAC-like methods. Process noise or initial condition noise in conjunction with a physics-based dynamics results in improved robustness on hard visual problems.
\n
\nA particle filter applied to the task of full 6D visual tracking of the pose an object being pushed by a robot in a table-top environment is improved on difficult visual problems by incorporating a simulator as a dynamics model and injecting noise as forces into the simulator.