Overcoming occlusions in eye-in-hand visual search

Abstract

In this paper we propose a method for handling persistent visual occlusions that disrupt visual tracking for eye-in-hand systems. This approach provides an efficient strategy for the robot to “look behind” the occlusion while respecting the robot's physical constraints. Specifically, we propose a decoupled search strategy combining a naïve pan tilt search with a sensor placement approach, to reduce the strategy's computational cost. We proceed by mapping limited environmental data into the robot configuration space and then planning within a constrained region. We use a particle filter to continuously estimate the target location, while our configuration-based cost function plans a goal location for the camera frame, taking into account robot singularity, self-collision and joint limit constraints. To validate our algorithm, we implemented it on an eye-in-hand robot system. Experimental results for various situations support the feasibility of our approach for quickly recovering fully occluded moving targets. Finally we discuss the implications of this approach to mobile robot platforms.