Exploiting the Trifocal Tensor in Dynamic Pose Estimation for Visual Control

摘要

Image-based approaches for visual control are memoryless, and they depend on the information extracted from the image plane. We propose the use of dynamic pose estimation in the task of driving a mobile robot to a desired location specified by a target image. This approach reduces the dependence of the control on the quality of current visual data and facilitates the planning of complex tasks. The pose estimation exploits the 1-D trifocal tensor (TT) as measurement, which allows us to obtain a semicalibrated estimation scheme that is valid for any visual sensor obeying a central projection model. The contribution of this brief is a novel observability analysis of the estimation problem from the 1-D TT using nonlinear tools, as well as the demonstration of the validity of closed-loop control from the estimated pose by showing a separation principle in our nonlinear framework. The overall position-based scheme drives the robot to a desired pose through smooth velocities without the need of a target model, either scene reconstruction or depth information. The effectiveness of the approach is evaluated via real-world experiments.