A Rigid Visual Servoing Approach for Elastic Objects Using Weighted Primitives

Abstract

Visual servoing on deformable objects is a recent and open problem. Many approaches propose to use continuous mechanical models embedded into the interaction matrix and compute corresponding camera visual servoing. These methods are highly parametric and are fine-tuned for every specific type of elastic objects. In this paper, we propose a generic approach that enables rigid camera visual servoing to fit as much as possible visual features of objects subjected to non-rigid deformation. The proposed method uses a triangular mesh representation of the visualized object. It computes a relative camera rigid repositioning for every triangular primitive. A specific proportional gain is used for every triangle. The proportional gain represents the relevance of every triangle in the global repositioning of the camera. This gain depends on the variation of the area of every triangle. This has the advantage of making the gain variation continuous which provides smooth camera repositioning. A global rigid correction of camera pose is computed as the scaled mean of all relative per-triangle rigid repositioning. The scale is taken to be the sum of all weights. We provide simulation results to show that with the weighted triangular primitives we are able to better fit the reference visual cue than if we used a naive approach of global uniform averaging of the per-triangle camera repositioning.