The evaluation and automation of tasks involving the manipulation of deformable curvilinear objects, such as threads and cables, requires the real-time estimation of the 3D shapes of these objects from images. This estimation is however extremely challenging due to the small amount of available visual information, the inherent geometric ambiguities, and the large object deformations.
We propose an approach for tracking deformable curvilinear objects using solely visual information from one or more calibrated cameras. The key idea is to formulate the shape estimation as a deformable 1D template tracking problem. The object is first textured with a pattern of different alternating colors. The tracking problem is then expressed as an energy minimization over a set of control points parameterizing a 3D NURBS modeling the object. Assuming the object’s in-extensibility, we propose a novel energy based on a texture-sensitive distance map.
We demonstrate the benefits of this energy in synthetic and real experiments, using data illustrating the deformation and manipulation of a thread with a da Vinci robot. In particular, we show that the approach allows for deformable tracking in the absence of normal motion along the curve, a challenging practical situation that occurs when the thread is dragged by one extremity.